{"schema_version":1,"snapshot_utc":"2026-11-10T20:57:03.676588+00:00","reference_cell":"cap_10000_n100","primary_analysis_view":"strict","models":{"AUTOGLUON":{"id":"AUTOGLUON","display":"AutoGluon","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"AutoGluon automates tabular learning by training multiple models and combining their predictions in ensembles. Here it runs its native extreme preset within a one-hour budget as a separate AutoML reference.","tuning":null,"category":"AutoML","color":"#111827","regular_max_features":null,"training_data_overlap":false},"CAT":{"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},"CAUSILO":{"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},"DUMMY":{"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},"FTFM":{"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},"GBM":{"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},"GENEICL":{"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},"KNN":{"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},"KUMO-TABULAR-LARGE":{"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},"KUMO-TABULAR-MEDIUM":{"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},"LIMIX-2":{"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},"LR":{"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},"LR-TUNED":{"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},"MITRA":{"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},"NN_TORCH":{"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},"NN_TORCH-TUNED":{"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},"REALMLP":{"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},"REALTABPFN-V2":{"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},"REALTABPFN-V2.5":{"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},"RF":{"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},"RF-TUNED":{"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},"SELDON":{"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},"SELDON-MINI":{"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},"TABDPT":{"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},"TABFM":{"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},"TABICL":{"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},"TABM":{"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},"TABPFN-V3":{"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},"TABPFN-V3.5":{"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},"TABPFN-WIDE":{"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},"TABPFN-WIDE-5K-NE3":{"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},"XGB":{"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},"XGB-TUNED":{"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},"XT":{"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},"XT-TUNED":{"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false}},"cell_options":[{"id":"cap_2000_n20","label":"p=2,000, n=20","feature_cap":2000,"n_train":20},{"id":"cap_2000_n50","label":"p=2,000, n=50","feature_cap":2000,"n_train":50},{"id":"cap_2000_n100","label":"p=2,000, n=100","feature_cap":2000,"n_train":100},{"id":"cap_2000_n200","label":"p=2,000, n=200","feature_cap":2000,"n_train":200},{"id":"cap_2000_n500","label":"p=2,000, n=500","feature_cap":2000,"n_train":500},{"id":"cap_2000_n1000","label":"p=2,000, n=1,000","feature_cap":2000,"n_train":1000},{"id":"cap_2000","label":"p=2,000, n=all","feature_cap":2000,"n_train":null},{"id":"cap_10000_n20","label":"p=10,000, n=20","feature_cap":10000,"n_train":20},{"id":"cap_10000_n50","label":"p=10,000, n=50","feature_cap":10000,"n_train":50},{"id":"cap_10000_n100","label":"p=10,000, n=100","feature_cap":10000,"n_train":100},{"id":"cap_10000_n200","label":"p=10,000, n=200","feature_cap":10000,"n_train":200},{"id":"cap_10000_n500","label":"p=10,000, n=500","feature_cap":10000,"n_train":500},{"id":"cap_10000_n1000","label":"p=10,000, n=1,000","feature_cap":10000,"n_train":1000},{"id":"cap_10000","label":"p=10,000, n=all","feature_cap":10000,"n_train":null},{"id":"cap_25000_n20","label":"p=25,000, n=20","feature_cap":25000,"n_train":20},{"id":"cap_25000_n50","label":"p=25,000, n=50","feature_cap":25000,"n_train":50},{"id":"cap_25000_n100","label":"p=25,000, n=100","feature_cap":25000,"n_train":100},{"id":"cap_25000_n200","label":"p=25,000, n=200","feature_cap":25000,"n_train":200},{"id":"cap_25000_n500","label":"p=25,000, n=500","feature_cap":25000,"n_train":500},{"id":"cap_25000_n1000","label":"p=25,000, n=1,000","feature_cap":25000,"n_train":1000},{"id":"cap_25000","label":"p=25,000, n=all","feature_cap":25000,"n_train":null},{"id":"cap_full_n20","label":"p=all, n=20","feature_cap":null,"n_train":20},{"id":"cap_full_n50","label":"p=all, n=50","feature_cap":null,"n_train":50},{"id":"cap_full_n100","label":"p=all, n=100","feature_cap":null,"n_train":100},{"id":"cap_full_n200","label":"p=all, n=200","feature_cap":null,"n_train":200},{"id":"cap_full_n500","label":"p=all, n=500","feature_cap":null,"n_train":500},{"id":"cap_full_n1000","label":"p=all, n=1,000","feature_cap":null,"n_train":1000},{"id":"cap_full","label":"p=all, n=all","feature_cap":null,"n_train":null}],"reference":[{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1239,"Elo_lo":1182,"Elo_hi":1302,"n_targets":44,"f1_macro":0.737513,"rmse":1.75962,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1226,"Elo_lo":1159,"Elo_hi":1290,"n_targets":44,"f1_macro":0.740375,"rmse":1.8159,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1210,"Elo_lo":1153,"Elo_hi":1268,"n_targets":32,"f1_macro":0.73223,"rmse":null,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1210,"Elo_lo":1145,"Elo_hi":1280,"n_targets":44,"f1_macro":0.719262,"rmse":1.7985,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1199,"Elo_lo":1136,"Elo_hi":1268,"n_targets":44,"f1_macro":0.713243,"rmse":1.85064,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1193,"Elo_lo":1151,"Elo_hi":1237,"n_targets":44,"f1_macro":0.719811,"rmse":1.80462,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1158,"Elo_lo":1107,"Elo_hi":1206,"n_targets":44,"f1_macro":0.702762,"rmse":1.74289,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1110,"Elo_lo":1032,"Elo_hi":1184,"n_targets":44,"f1_macro":0.672508,"rmse":1.83574,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1072,"Elo_lo":1024,"Elo_hi":1119,"n_targets":44,"f1_macro":0.673781,"rmse":1.83726,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1063,"Elo_lo":998,"Elo_hi":1125,"n_targets":44,"f1_macro":0.707116,"rmse":1.95582,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1060,"Elo_lo":1001,"Elo_hi":1120,"n_targets":43,"f1_macro":0.685468,"rmse":1.89663,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1055,"Elo_lo":1002,"Elo_hi":1113,"n_targets":44,"f1_macro":0.685489,"rmse":1.96288,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1050,"Elo_lo":975,"Elo_hi":1120,"n_targets":44,"f1_macro":0.692084,"rmse":1.92319,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1046,"Elo_lo":993,"Elo_hi":1098,"n_targets":44,"f1_macro":0.654833,"rmse":1.8178,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"AUTOGLUON","Elo":1040,"Elo_lo":992,"Elo_hi":1090,"n_targets":44,"f1_macro":0.684564,"rmse":2.07258,"id":"AUTOGLUON","display":"AutoGluon","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"AutoGluon automates tabular learning by training multiple models and combining their predictions in ensembles. Here it runs its native extreme preset within a one-hour budget as a separate AutoML reference.","tuning":null,"category":"AutoML","color":"#111827","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1040,"Elo_lo":984,"Elo_hi":1095,"n_targets":44,"f1_macro":0.686295,"rmse":2.11554,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1040,"Elo_lo":990,"Elo_hi":1092,"n_targets":44,"f1_macro":0.684965,"rmse":2.09472,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1039,"Elo_lo":988,"Elo_hi":1094,"n_targets":44,"f1_macro":0.683193,"rmse":2.12524,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1035,"Elo_lo":958,"Elo_hi":1103,"n_targets":32,"f1_macro":0.662519,"rmse":null,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1023,"Elo_lo":942,"Elo_hi":1094,"n_targets":32,"f1_macro":0.655745,"rmse":null,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1018,"Elo_lo":954,"Elo_hi":1085,"n_targets":44,"f1_macro":0.629495,"rmse":1.79114,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1014,"Elo_lo":972,"Elo_hi":1051,"n_targets":44,"f1_macro":0.662073,"rmse":2.02183,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1011,"Elo_lo":969,"Elo_hi":1061,"n_targets":44,"f1_macro":0.668661,"rmse":2.01502,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1008,"Elo_lo":940,"Elo_hi":1071,"n_targets":44,"f1_macro":0.689285,"rmse":2.17055,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"f1_macro":0.65413,"rmse":1.93045,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":991,"Elo_lo":962,"Elo_hi":1023,"n_targets":44,"f1_macro":0.647685,"rmse":1.99831,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":980,"Elo_lo":946,"Elo_hi":1017,"n_targets":44,"f1_macro":0.63455,"rmse":1.97764,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT","Elo":976,"Elo_lo":952,"Elo_hi":998,"n_targets":44,"f1_macro":0.633911,"rmse":1.99815,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":967,"Elo_lo":918,"Elo_hi":1016,"n_targets":44,"f1_macro":0.636697,"rmse":1.90252,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":915,"Elo_lo":860,"Elo_hi":966,"n_targets":44,"f1_macro":0.628504,"rmse":1.88066,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":909,"Elo_lo":863,"Elo_hi":954,"n_targets":44,"f1_macro":0.613568,"rmse":1.92813,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":880,"Elo_lo":832,"Elo_hi":927,"n_targets":44,"f1_macro":0.625905,"rmse":2.10961,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":873,"Elo_lo":821,"Elo_hi":924,"n_targets":44,"f1_macro":0.619756,"rmse":2.18334,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":788,"Elo_lo":687,"Elo_hi":876,"n_targets":44,"f1_macro":0.371345,"rmse":2.16769,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":558,"Elo_lo":487,"Elo_hi":616,"n_targets":44,"f1_macro":0.33199,"rmse":2.23666,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false}],"cell_elo":[{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1153,"Elo_lo":1102,"Elo_hi":1208,"n_targets":45,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1144,"Elo_lo":1087,"Elo_hi":1207,"n_targets":45,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1120,"Elo_lo":1062,"Elo_hi":1183,"n_targets":45,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1116,"Elo_lo":1057,"Elo_hi":1174,"n_targets":45,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1108,"Elo_lo":1061,"Elo_hi":1153,"n_targets":45,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1075,"Elo_lo":1032,"Elo_hi":1122,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1068,"Elo_lo":1021,"Elo_hi":1114,"n_targets":45,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1067,"Elo_lo":1022,"Elo_hi":1109,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1041,"Elo_lo":987,"Elo_hi":1091,"n_targets":45,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1039,"Elo_lo":995,"Elo_hi":1082,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1024,"Elo_lo":976,"Elo_hi":1069,"n_targets":45,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1021,"Elo_lo":980,"Elo_hi":1063,"n_targets":45,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1005,"Elo_lo":942,"Elo_hi":1065,"n_targets":44,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1002,"Elo_lo":940,"Elo_hi":1066,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT","Elo":1001,"Elo_lo":975,"Elo_hi":1024,"n_targets":45,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":45,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":995,"Elo_lo":957,"Elo_hi":1025,"n_targets":45,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":991,"Elo_lo":926,"Elo_hi":1060,"n_targets":45,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":968,"Elo_lo":905,"Elo_hi":1030,"n_targets":45,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":963,"Elo_lo":897,"Elo_hi":1030,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":951,"Elo_lo":897,"Elo_hi":1005,"n_targets":45,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":946,"Elo_lo":898,"Elo_hi":994,"n_targets":45,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":936,"Elo_lo":882,"Elo_hi":988,"n_targets":45,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":928,"Elo_lo":881,"Elo_hi":969,"n_targets":45,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":920,"Elo_lo":867,"Elo_hi":963,"n_targets":45,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":910,"Elo_lo":832,"Elo_hi":985,"n_targets":45,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":902,"Elo_lo":842,"Elo_hi":962,"n_targets":45,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":900,"Elo_lo":840,"Elo_hi":958,"n_targets":45,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":891,"Elo_lo":843,"Elo_hi":933,"n_targets":45,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":890,"Elo_lo":838,"Elo_hi":943,"n_targets":45,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":845,"Elo_lo":799,"Elo_hi":887,"n_targets":45,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":805,"Elo_lo":754,"Elo_hi":853,"n_targets":45,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":740,"Elo_lo":667,"Elo_hi":807,"n_targets":45,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":739,"Elo_lo":666,"Elo_hi":808,"n_targets":45,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1239,"Elo_lo":1177,"Elo_hi":1303,"n_targets":46,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1232,"Elo_lo":1179,"Elo_hi":1289,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1228,"Elo_lo":1175,"Elo_hi":1278,"n_targets":46,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1222,"Elo_lo":1161,"Elo_hi":1283,"n_targets":46,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1212,"Elo_lo":1154,"Elo_hi":1270,"n_targets":46,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1143,"Elo_lo":1097,"Elo_hi":1191,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1122,"Elo_lo":1073,"Elo_hi":1170,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1103,"Elo_lo":1060,"Elo_hi":1151,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1090,"Elo_lo":1014,"Elo_hi":1164,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1086,"Elo_lo":1038,"Elo_hi":1133,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1073,"Elo_lo":1024,"Elo_hi":1120,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1069,"Elo_lo":1022,"Elo_hi":1118,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1061,"Elo_lo":1010,"Elo_hi":1108,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1027,"Elo_lo":980,"Elo_hi":1075,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1023,"Elo_lo":972,"Elo_hi":1074,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1021,"Elo_lo":974,"Elo_hi":1066,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1020,"Elo_lo":946,"Elo_hi":1093,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1009,"Elo_lo":943,"Elo_hi":1080,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1001,"Elo_lo":956,"Elo_hi":1044,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":999,"Elo_lo":944,"Elo_hi":1056,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT","Elo":996,"Elo_lo":972,"Elo_hi":1022,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":984,"Elo_lo":938,"Elo_hi":1028,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":983,"Elo_lo":955,"Elo_hi":1010,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":980,"Elo_lo":930,"Elo_hi":1031,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":979,"Elo_lo":915,"Elo_hi":1038,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":974,"Elo_lo":940,"Elo_hi":1007,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":960,"Elo_lo":929,"Elo_hi":992,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":928,"Elo_lo":879,"Elo_hi":977,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":869,"Elo_lo":793,"Elo_hi":944,"n_targets":46,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":845,"Elo_lo":797,"Elo_hi":891,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":838,"Elo_lo":802,"Elo_hi":872,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":790,"Elo_lo":733,"Elo_hi":842,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":651,"Elo_lo":578,"Elo_hi":719,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1276,"Elo_lo":1212,"Elo_hi":1347,"n_targets":44,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1256,"Elo_lo":1194,"Elo_hi":1321,"n_targets":44,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1255,"Elo_lo":1184,"Elo_hi":1328,"n_targets":44,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1245,"Elo_lo":1169,"Elo_hi":1320,"n_targets":44,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1229,"Elo_lo":1160,"Elo_hi":1300,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1215,"Elo_lo":1171,"Elo_hi":1265,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1153,"Elo_lo":1096,"Elo_hi":1213,"n_targets":44,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1143,"Elo_lo":1069,"Elo_hi":1213,"n_targets":44,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1106,"Elo_lo":1053,"Elo_hi":1162,"n_targets":44,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1105,"Elo_lo":1035,"Elo_hi":1175,"n_targets":44,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1100,"Elo_lo":1040,"Elo_hi":1161,"n_targets":44,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1098,"Elo_lo":1043,"Elo_hi":1156,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1083,"Elo_lo":1019,"Elo_hi":1145,"n_targets":44,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1078,"Elo_lo":1000,"Elo_hi":1151,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1078,"Elo_lo":1019,"Elo_hi":1140,"n_targets":44,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1074,"Elo_lo":1016,"Elo_hi":1131,"n_targets":43,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1074,"Elo_lo":1022,"Elo_hi":1130,"n_targets":44,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1055,"Elo_lo":972,"Elo_hi":1134,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1044,"Elo_lo":993,"Elo_hi":1094,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1044,"Elo_lo":1000,"Elo_hi":1092,"n_targets":44,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1037,"Elo_lo":980,"Elo_hi":1095,"n_targets":44,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1034,"Elo_lo":955,"Elo_hi":1106,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1020,"Elo_lo":984,"Elo_hi":1058,"n_targets":44,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1012,"Elo_lo":940,"Elo_hi":1076,"n_targets":44,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1004,"Elo_lo":979,"Elo_hi":1027,"n_targets":44,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT","Elo":988,"Elo_lo":968,"Elo_hi":1006,"n_targets":44,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":977,"Elo_lo":947,"Elo_hi":1010,"n_targets":44,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":917,"Elo_lo":862,"Elo_hi":967,"n_targets":44,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":908,"Elo_lo":864,"Elo_hi":953,"n_targets":44,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":887,"Elo_lo":836,"Elo_hi":938,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":873,"Elo_lo":810,"Elo_hi":937,"n_targets":44,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":862,"Elo_lo":774,"Elo_hi":943,"n_targets":44,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":581,"Elo_lo":504,"Elo_hi":644,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1283,"Elo_lo":1210,"Elo_hi":1359,"n_targets":36,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1228,"Elo_lo":1156,"Elo_hi":1309,"n_targets":36,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1215,"Elo_lo":1137,"Elo_hi":1297,"n_targets":36,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1213,"Elo_lo":1136,"Elo_hi":1292,"n_targets":36,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1212,"Elo_lo":1153,"Elo_hi":1278,"n_targets":36,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1195,"Elo_lo":1128,"Elo_hi":1265,"n_targets":36,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1179,"Elo_lo":1106,"Elo_hi":1259,"n_targets":24,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1177,"Elo_lo":1113,"Elo_hi":1246,"n_targets":36,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1131,"Elo_lo":1048,"Elo_hi":1220,"n_targets":36,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1127,"Elo_lo":1062,"Elo_hi":1195,"n_targets":36,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1111,"Elo_lo":1042,"Elo_hi":1188,"n_targets":36,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1108,"Elo_lo":1022,"Elo_hi":1198,"n_targets":36,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1108,"Elo_lo":1033,"Elo_hi":1185,"n_targets":36,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1095,"Elo_lo":1026,"Elo_hi":1164,"n_targets":36,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1088,"Elo_lo":1000,"Elo_hi":1176,"n_targets":24,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1081,"Elo_lo":995,"Elo_hi":1160,"n_targets":24,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1077,"Elo_lo":1018,"Elo_hi":1136,"n_targets":36,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1071,"Elo_lo":1007,"Elo_hi":1141,"n_targets":36,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1064,"Elo_lo":994,"Elo_hi":1132,"n_targets":36,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1059,"Elo_lo":987,"Elo_hi":1127,"n_targets":36,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1041,"Elo_lo":953,"Elo_hi":1124,"n_targets":36,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1037,"Elo_lo":996,"Elo_hi":1078,"n_targets":36,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1019,"Elo_lo":953,"Elo_hi":1086,"n_targets":36,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1014,"Elo_lo":956,"Elo_hi":1082,"n_targets":36,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1004,"Elo_lo":968,"Elo_hi":1039,"n_targets":36,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":36,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT","Elo":989,"Elo_lo":970,"Elo_hi":1007,"n_targets":36,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":988,"Elo_lo":940,"Elo_hi":1038,"n_targets":36,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":959,"Elo_lo":904,"Elo_hi":1014,"n_targets":36,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":949,"Elo_lo":908,"Elo_hi":992,"n_targets":36,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":921,"Elo_lo":810,"Elo_hi":1020,"n_targets":36,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":916,"Elo_lo":861,"Elo_hi":973,"n_targets":36,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":841,"Elo_lo":768,"Elo_hi":911,"n_targets":36,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":496,"Elo_lo":422,"Elo_hi":569,"n_targets":36,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1231,"Elo_lo":1132,"Elo_hi":1348,"n_targets":22,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1211,"Elo_lo":1124,"Elo_hi":1308,"n_targets":22,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1189,"Elo_lo":1091,"Elo_hi":1292,"n_targets":22,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1181,"Elo_lo":1073,"Elo_hi":1301,"n_targets":22,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1155,"Elo_lo":1045,"Elo_hi":1268,"n_targets":22,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1141,"Elo_lo":1032,"Elo_hi":1260,"n_targets":22,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1133,"Elo_lo":1044,"Elo_hi":1224,"n_targets":22,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1111,"Elo_lo":1027,"Elo_hi":1203,"n_targets":22,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1099,"Elo_lo":993,"Elo_hi":1222,"n_targets":14,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1098,"Elo_lo":994,"Elo_hi":1218,"n_targets":14,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1069,"Elo_lo":1001,"Elo_hi":1143,"n_targets":22,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1056,"Elo_lo":955,"Elo_hi":1162,"n_targets":22,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1041,"Elo_lo":965,"Elo_hi":1120,"n_targets":22,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1039,"Elo_lo":949,"Elo_hi":1123,"n_targets":22,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1036,"Elo_lo":902,"Elo_hi":1181,"n_targets":22,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1027,"Elo_lo":943,"Elo_hi":1113,"n_targets":22,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1021,"Elo_lo":941,"Elo_hi":1104,"n_targets":22,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1016,"Elo_lo":919,"Elo_hi":1106,"n_targets":22,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1011,"Elo_lo":881,"Elo_hi":1146,"n_targets":14,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1010,"Elo_lo":960,"Elo_hi":1059,"n_targets":22,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1008,"Elo_lo":867,"Elo_hi":1149,"n_targets":22,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1006,"Elo_lo":977,"Elo_hi":1035,"n_targets":22,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT","Elo":1003,"Elo_lo":969,"Elo_hi":1036,"n_targets":22,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":22,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":986,"Elo_lo":902,"Elo_hi":1071,"n_targets":22,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":978,"Elo_lo":922,"Elo_hi":1033,"n_targets":22,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":977,"Elo_lo":919,"Elo_hi":1032,"n_targets":22,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":954,"Elo_lo":895,"Elo_hi":1021,"n_targets":22,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":916,"Elo_lo":819,"Elo_hi":1013,"n_targets":22,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR","Elo":901,"Elo_lo":765,"Elo_hi":1015,"n_targets":22,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":897,"Elo_lo":845,"Elo_hi":959,"n_targets":22,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":817,"Elo_lo":654,"Elo_hi":950,"n_targets":22,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":762,"Elo_lo":660,"Elo_hi":860,"n_targets":22,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":411,"Elo_lo":310,"Elo_hi":519,"n_targets":22,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1202,"Elo_lo":1066,"Elo_hi":1358,"n_targets":17,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1196,"Elo_lo":1089,"Elo_hi":1310,"n_targets":17,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1183,"Elo_lo":1057,"Elo_hi":1310,"n_targets":17,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1177,"Elo_lo":1064,"Elo_hi":1305,"n_targets":17,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1160,"Elo_lo":1025,"Elo_hi":1296,"n_targets":17,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1099,"Elo_lo":945,"Elo_hi":1265,"n_targets":17,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1092,"Elo_lo":969,"Elo_hi":1211,"n_targets":10,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1092,"Elo_lo":943,"Elo_hi":1244,"n_targets":17,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1085,"Elo_lo":938,"Elo_hi":1233,"n_targets":10,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1077,"Elo_lo":954,"Elo_hi":1222,"n_targets":17,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1038,"Elo_lo":910,"Elo_hi":1165,"n_targets":17,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1026,"Elo_lo":918,"Elo_hi":1142,"n_targets":17,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1011,"Elo_lo":929,"Elo_hi":1091,"n_targets":17,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":17,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":997,"Elo_lo":858,"Elo_hi":1141,"n_targets":17,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":990,"Elo_lo":919,"Elo_hi":1057,"n_targets":17,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT","Elo":990,"Elo_lo":957,"Elo_hi":1021,"n_targets":17,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":988,"Elo_lo":850,"Elo_hi":1111,"n_targets":17,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":976,"Elo_lo":839,"Elo_hi":1099,"n_targets":17,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":967,"Elo_lo":917,"Elo_hi":1018,"n_targets":17,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":966,"Elo_lo":801,"Elo_hi":1151,"n_targets":17,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":956,"Elo_lo":882,"Elo_hi":1020,"n_targets":17,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":946,"Elo_lo":843,"Elo_hi":1039,"n_targets":17,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":930,"Elo_lo":778,"Elo_hi":1067,"n_targets":17,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":930,"Elo_lo":836,"Elo_hi":1013,"n_targets":17,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":925,"Elo_lo":722,"Elo_hi":1123,"n_targets":17,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":923,"Elo_lo":746,"Elo_hi":1102,"n_targets":10,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":854,"Elo_lo":705,"Elo_hi":985,"n_targets":17,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":838,"Elo_lo":710,"Elo_hi":977,"n_targets":17,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR","Elo":777,"Elo_lo":533,"Elo_hi":971,"n_targets":17,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":737,"Elo_lo":602,"Elo_hi":861,"n_targets":17,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":709,"Elo_lo":588,"Elo_hi":811,"n_targets":17,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":525,"Elo_lo":247,"Elo_hi":757,"n_targets":17,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":364,"Elo_lo":198,"Elo_hi":507,"n_targets":17,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1249,"Elo_lo":1189,"Elo_hi":1309,"n_targets":46,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1240,"Elo_lo":1184,"Elo_hi":1297,"n_targets":46,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1235,"Elo_lo":1174,"Elo_hi":1297,"n_targets":46,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1221,"Elo_lo":1173,"Elo_hi":1270,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1215,"Elo_lo":1155,"Elo_hi":1274,"n_targets":46,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1171,"Elo_lo":1110,"Elo_hi":1235,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1158,"Elo_lo":1091,"Elo_hi":1230,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1154,"Elo_lo":1077,"Elo_hi":1231,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1135,"Elo_lo":1079,"Elo_hi":1196,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1114,"Elo_lo":1061,"Elo_hi":1171,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1113,"Elo_lo":1060,"Elo_hi":1171,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1101,"Elo_lo":1034,"Elo_hi":1171,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1093,"Elo_lo":1048,"Elo_hi":1140,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1080,"Elo_lo":1021,"Elo_hi":1137,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1074,"Elo_lo":1018,"Elo_hi":1127,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1073,"Elo_lo":1016,"Elo_hi":1128,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1070,"Elo_lo":1007,"Elo_hi":1133,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1052,"Elo_lo":1006,"Elo_hi":1099,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1045,"Elo_lo":950,"Elo_hi":1137,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1042,"Elo_lo":1006,"Elo_hi":1076,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1040,"Elo_lo":984,"Elo_hi":1096,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1033,"Elo_lo":961,"Elo_hi":1100,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1029,"Elo_lo":1000,"Elo_hi":1055,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1013,"Elo_lo":966,"Elo_hi":1055,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1003,"Elo_lo":981,"Elo_hi":1026,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR","Elo":995,"Elo_lo":919,"Elo_hi":1062,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT","Elo":992,"Elo_lo":970,"Elo_hi":1013,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":987,"Elo_lo":911,"Elo_hi":1057,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":987,"Elo_lo":944,"Elo_hi":1028,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":976,"Elo_lo":906,"Elo_hi":1039,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":881,"Elo_lo":835,"Elo_hi":929,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":648,"Elo_lo":516,"Elo_hi":761,"n_targets":46,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":483,"Elo_lo":415,"Elo_hi":544,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1147,"Elo_lo":1098,"Elo_hi":1200,"n_targets":45,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1142,"Elo_lo":1081,"Elo_hi":1204,"n_targets":45,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1117,"Elo_lo":1053,"Elo_hi":1185,"n_targets":45,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1098,"Elo_lo":1054,"Elo_hi":1143,"n_targets":45,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1084,"Elo_lo":1026,"Elo_hi":1141,"n_targets":45,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1075,"Elo_lo":1034,"Elo_hi":1120,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1069,"Elo_lo":1021,"Elo_hi":1114,"n_targets":45,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1060,"Elo_lo":1016,"Elo_hi":1103,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1053,"Elo_lo":999,"Elo_hi":1105,"n_targets":45,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1048,"Elo_lo":1006,"Elo_hi":1089,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1009,"Elo_lo":945,"Elo_hi":1073,"n_targets":44,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1006,"Elo_lo":948,"Elo_hi":1067,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT","Elo":1005,"Elo_lo":987,"Elo_hi":1025,"n_targets":45,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1003,"Elo_lo":955,"Elo_hi":1048,"n_targets":45,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":45,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":997,"Elo_lo":969,"Elo_hi":1024,"n_targets":45,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":983,"Elo_lo":918,"Elo_hi":1049,"n_targets":45,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":977,"Elo_lo":920,"Elo_hi":1034,"n_targets":45,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":971,"Elo_lo":909,"Elo_hi":1032,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":969,"Elo_lo":932,"Elo_hi":1010,"n_targets":45,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":964,"Elo_lo":904,"Elo_hi":1020,"n_targets":45,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":958,"Elo_lo":918,"Elo_hi":997,"n_targets":45,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":947,"Elo_lo":899,"Elo_hi":994,"n_targets":45,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":940,"Elo_lo":883,"Elo_hi":995,"n_targets":45,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":923,"Elo_lo":878,"Elo_hi":963,"n_targets":45,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":918,"Elo_lo":870,"Elo_hi":964,"n_targets":45,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":911,"Elo_lo":853,"Elo_hi":969,"n_targets":45,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":906,"Elo_lo":862,"Elo_hi":948,"n_targets":45,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":899,"Elo_lo":851,"Elo_hi":950,"n_targets":45,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":846,"Elo_lo":805,"Elo_hi":886,"n_targets":45,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":808,"Elo_lo":724,"Elo_hi":888,"n_targets":45,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":802,"Elo_lo":752,"Elo_hi":847,"n_targets":45,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":738,"Elo_lo":668,"Elo_hi":801,"n_targets":45,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":734,"Elo_lo":665,"Elo_hi":801,"n_targets":45,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1257,"Elo_lo":1201,"Elo_hi":1316,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1222,"Elo_lo":1169,"Elo_hi":1272,"n_targets":46,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1218,"Elo_lo":1163,"Elo_hi":1274,"n_targets":46,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1213,"Elo_lo":1157,"Elo_hi":1271,"n_targets":46,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1175,"Elo_lo":1122,"Elo_hi":1231,"n_targets":46,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1142,"Elo_lo":1095,"Elo_hi":1187,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1128,"Elo_lo":1083,"Elo_hi":1177,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1087,"Elo_lo":1037,"Elo_hi":1139,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1085,"Elo_lo":1036,"Elo_hi":1134,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1078,"Elo_lo":1029,"Elo_hi":1122,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1065,"Elo_lo":1017,"Elo_hi":1114,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1048,"Elo_lo":977,"Elo_hi":1126,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1034,"Elo_lo":958,"Elo_hi":1112,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1027,"Elo_lo":978,"Elo_hi":1079,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1024,"Elo_lo":974,"Elo_hi":1074,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1016,"Elo_lo":964,"Elo_hi":1069,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1014,"Elo_lo":947,"Elo_hi":1081,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1013,"Elo_lo":959,"Elo_hi":1068,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1013,"Elo_lo":954,"Elo_hi":1071,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1000,"Elo_lo":937,"Elo_hi":1062,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT","Elo":994,"Elo_lo":972,"Elo_hi":1016,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":991,"Elo_lo":937,"Elo_hi":1043,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":990,"Elo_lo":949,"Elo_hi":1033,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":973,"Elo_lo":942,"Elo_hi":1007,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":965,"Elo_lo":924,"Elo_hi":1004,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":964,"Elo_lo":929,"Elo_hi":1002,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":962,"Elo_lo":914,"Elo_hi":1011,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":938,"Elo_lo":886,"Elo_hi":993,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":839,"Elo_lo":800,"Elo_hi":877,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":838,"Elo_lo":798,"Elo_hi":876,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":828,"Elo_lo":778,"Elo_hi":874,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":804,"Elo_lo":728,"Elo_hi":882,"n_targets":46,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":646,"Elo_lo":578,"Elo_hi":708,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1239,"Elo_lo":1182,"Elo_hi":1302,"n_targets":44,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1226,"Elo_lo":1159,"Elo_hi":1290,"n_targets":44,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1210,"Elo_lo":1153,"Elo_hi":1268,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1210,"Elo_lo":1145,"Elo_hi":1280,"n_targets":44,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1199,"Elo_lo":1136,"Elo_hi":1268,"n_targets":44,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1193,"Elo_lo":1151,"Elo_hi":1237,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1158,"Elo_lo":1107,"Elo_hi":1206,"n_targets":44,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1110,"Elo_lo":1032,"Elo_hi":1184,"n_targets":44,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1072,"Elo_lo":1024,"Elo_hi":1119,"n_targets":44,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1063,"Elo_lo":998,"Elo_hi":1125,"n_targets":44,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1060,"Elo_lo":1001,"Elo_hi":1120,"n_targets":43,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1055,"Elo_lo":1002,"Elo_hi":1113,"n_targets":44,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1050,"Elo_lo":975,"Elo_hi":1120,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1046,"Elo_lo":993,"Elo_hi":1098,"n_targets":44,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"AUTOGLUON","Elo":1040,"Elo_lo":992,"Elo_hi":1090,"n_targets":44,"id":"AUTOGLUON","display":"AutoGluon","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"AutoGluon automates tabular learning by training multiple models and combining their predictions in ensembles. Here it runs its native extreme preset within a one-hour budget as a separate AutoML reference.","tuning":null,"category":"AutoML","color":"#111827","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1040,"Elo_lo":984,"Elo_hi":1095,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1040,"Elo_lo":990,"Elo_hi":1092,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1039,"Elo_lo":988,"Elo_hi":1094,"n_targets":44,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1035,"Elo_lo":958,"Elo_hi":1103,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1023,"Elo_lo":942,"Elo_hi":1094,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1018,"Elo_lo":954,"Elo_hi":1085,"n_targets":44,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1014,"Elo_lo":972,"Elo_hi":1051,"n_targets":44,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1011,"Elo_lo":969,"Elo_hi":1061,"n_targets":44,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1008,"Elo_lo":940,"Elo_hi":1071,"n_targets":44,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":991,"Elo_lo":962,"Elo_hi":1023,"n_targets":44,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":980,"Elo_lo":946,"Elo_hi":1017,"n_targets":44,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT","Elo":976,"Elo_lo":952,"Elo_hi":998,"n_targets":44,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":967,"Elo_lo":918,"Elo_hi":1016,"n_targets":44,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":915,"Elo_lo":860,"Elo_hi":966,"n_targets":44,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":909,"Elo_lo":863,"Elo_hi":954,"n_targets":44,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":880,"Elo_lo":832,"Elo_hi":927,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":873,"Elo_lo":821,"Elo_hi":924,"n_targets":44,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":788,"Elo_lo":687,"Elo_hi":876,"n_targets":44,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":558,"Elo_lo":487,"Elo_hi":616,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1261,"Elo_lo":1199,"Elo_hi":1329,"n_targets":36,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1221,"Elo_lo":1169,"Elo_hi":1276,"n_targets":36,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1211,"Elo_lo":1137,"Elo_hi":1288,"n_targets":36,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1211,"Elo_lo":1146,"Elo_hi":1278,"n_targets":36,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1204,"Elo_lo":1134,"Elo_hi":1269,"n_targets":36,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1163,"Elo_lo":1093,"Elo_hi":1241,"n_targets":24,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1162,"Elo_lo":1109,"Elo_hi":1225,"n_targets":36,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1158,"Elo_lo":1077,"Elo_hi":1243,"n_targets":36,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1120,"Elo_lo":1062,"Elo_hi":1185,"n_targets":36,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1101,"Elo_lo":1044,"Elo_hi":1157,"n_targets":36,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1100,"Elo_lo":1010,"Elo_hi":1188,"n_targets":24,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1080,"Elo_lo":999,"Elo_hi":1159,"n_targets":36,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1079,"Elo_lo":1011,"Elo_hi":1153,"n_targets":36,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1078,"Elo_lo":1000,"Elo_hi":1160,"n_targets":36,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1066,"Elo_lo":1003,"Elo_hi":1129,"n_targets":36,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1061,"Elo_lo":995,"Elo_hi":1130,"n_targets":36,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1061,"Elo_lo":996,"Elo_hi":1123,"n_targets":36,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1058,"Elo_lo":980,"Elo_hi":1134,"n_targets":36,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1055,"Elo_lo":965,"Elo_hi":1146,"n_targets":24,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1032,"Elo_lo":999,"Elo_hi":1066,"n_targets":36,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1022,"Elo_lo":960,"Elo_hi":1079,"n_targets":36,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1017,"Elo_lo":977,"Elo_hi":1061,"n_targets":36,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1009,"Elo_lo":971,"Elo_hi":1053,"n_targets":36,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1002,"Elo_lo":945,"Elo_hi":1066,"n_targets":36,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":36,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":996,"Elo_lo":926,"Elo_hi":1065,"n_targets":36,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":986,"Elo_lo":936,"Elo_hi":1041,"n_targets":36,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":983,"Elo_lo":931,"Elo_hi":1038,"n_targets":36,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT","Elo":982,"Elo_lo":959,"Elo_hi":1002,"n_targets":36,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":966,"Elo_lo":899,"Elo_hi":1030,"n_targets":36,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":930,"Elo_lo":871,"Elo_hi":996,"n_targets":36,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":870,"Elo_lo":752,"Elo_hi":974,"n_targets":36,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":845,"Elo_lo":785,"Elo_hi":901,"n_targets":36,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":499,"Elo_lo":428,"Elo_hi":569,"n_targets":36,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1213,"Elo_lo":1115,"Elo_hi":1323,"n_targets":22,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1183,"Elo_lo":1112,"Elo_hi":1267,"n_targets":22,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1157,"Elo_lo":1051,"Elo_hi":1266,"n_targets":22,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1157,"Elo_lo":1054,"Elo_hi":1260,"n_targets":22,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1149,"Elo_lo":1052,"Elo_hi":1249,"n_targets":22,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1137,"Elo_lo":1011,"Elo_hi":1267,"n_targets":14,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1114,"Elo_lo":1027,"Elo_hi":1203,"n_targets":22,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1104,"Elo_lo":1001,"Elo_hi":1214,"n_targets":22,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1088,"Elo_lo":1013,"Elo_hi":1173,"n_targets":22,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1087,"Elo_lo":975,"Elo_hi":1205,"n_targets":14,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1060,"Elo_lo":985,"Elo_hi":1138,"n_targets":22,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1048,"Elo_lo":953,"Elo_hi":1134,"n_targets":22,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1044,"Elo_lo":944,"Elo_hi":1148,"n_targets":22,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1044,"Elo_lo":964,"Elo_hi":1124,"n_targets":22,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1031,"Elo_lo":968,"Elo_hi":1086,"n_targets":22,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1029,"Elo_lo":994,"Elo_hi":1066,"n_targets":22,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1024,"Elo_lo":930,"Elo_hi":1110,"n_targets":22,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1014,"Elo_lo":922,"Elo_hi":1102,"n_targets":22,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1013,"Elo_lo":932,"Elo_hi":1089,"n_targets":22,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1004,"Elo_lo":861,"Elo_hi":1138,"n_targets":22,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":22,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT","Elo":999,"Elo_lo":972,"Elo_hi":1023,"n_targets":22,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":994,"Elo_lo":918,"Elo_hi":1064,"n_targets":22,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":987,"Elo_lo":859,"Elo_hi":1114,"n_targets":14,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":977,"Elo_lo":920,"Elo_hi":1037,"n_targets":22,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":976,"Elo_lo":886,"Elo_hi":1060,"n_targets":22,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR","Elo":945,"Elo_lo":798,"Elo_hi":1068,"n_targets":22,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":939,"Elo_lo":840,"Elo_hi":1032,"n_targets":22,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":928,"Elo_lo":857,"Elo_hi":992,"n_targets":22,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":885,"Elo_lo":793,"Elo_hi":972,"n_targets":22,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":827,"Elo_lo":699,"Elo_hi":958,"n_targets":22,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":801,"Elo_lo":638,"Elo_hi":935,"n_targets":22,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":758,"Elo_lo":651,"Elo_hi":848,"n_targets":22,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":433,"Elo_lo":331,"Elo_hi":538,"n_targets":22,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1201,"Elo_lo":1075,"Elo_hi":1343,"n_targets":17,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1193,"Elo_lo":1118,"Elo_hi":1291,"n_targets":17,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1167,"Elo_lo":1053,"Elo_hi":1280,"n_targets":17,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1127,"Elo_lo":992,"Elo_hi":1269,"n_targets":17,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1121,"Elo_lo":1024,"Elo_hi":1234,"n_targets":17,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1113,"Elo_lo":992,"Elo_hi":1234,"n_targets":10,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1108,"Elo_lo":970,"Elo_hi":1259,"n_targets":17,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1079,"Elo_lo":968,"Elo_hi":1212,"n_targets":17,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1076,"Elo_lo":955,"Elo_hi":1214,"n_targets":17,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1063,"Elo_lo":908,"Elo_hi":1226,"n_targets":10,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1041,"Elo_lo":937,"Elo_hi":1148,"n_targets":17,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1023,"Elo_lo":907,"Elo_hi":1137,"n_targets":17,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT","Elo":1003,"Elo_lo":969,"Elo_hi":1040,"n_targets":17,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":17,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":996,"Elo_lo":911,"Elo_hi":1082,"n_targets":17,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":993,"Elo_lo":953,"Elo_hi":1030,"n_targets":17,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":990,"Elo_lo":915,"Elo_hi":1067,"n_targets":17,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":990,"Elo_lo":864,"Elo_hi":1117,"n_targets":17,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":989,"Elo_lo":845,"Elo_hi":1134,"n_targets":17,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":987,"Elo_lo":932,"Elo_hi":1036,"n_targets":17,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":980,"Elo_lo":897,"Elo_hi":1053,"n_targets":17,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":975,"Elo_lo":842,"Elo_hi":1112,"n_targets":17,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":963,"Elo_lo":823,"Elo_hi":1100,"n_targets":17,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":934,"Elo_lo":850,"Elo_hi":1016,"n_targets":17,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":923,"Elo_lo":728,"Elo_hi":1110,"n_targets":17,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":901,"Elo_lo":710,"Elo_hi":1092,"n_targets":10,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":862,"Elo_lo":741,"Elo_hi":990,"n_targets":17,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":857,"Elo_lo":726,"Elo_hi":1014,"n_targets":17,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":844,"Elo_lo":692,"Elo_hi":977,"n_targets":17,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR","Elo":803,"Elo_lo":554,"Elo_hi":999,"n_targets":17,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":707,"Elo_lo":589,"Elo_hi":811,"n_targets":17,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":697,"Elo_lo":570,"Elo_hi":814,"n_targets":17,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":533,"Elo_lo":278,"Elo_hi":758,"n_targets":17,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":386,"Elo_lo":233,"Elo_hi":518,"n_targets":17,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1231,"Elo_lo":1167,"Elo_hi":1294,"n_targets":46,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1224,"Elo_lo":1160,"Elo_hi":1292,"n_targets":46,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1206,"Elo_lo":1161,"Elo_hi":1252,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1198,"Elo_lo":1136,"Elo_hi":1260,"n_targets":46,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1190,"Elo_lo":1131,"Elo_hi":1252,"n_targets":46,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1176,"Elo_lo":1116,"Elo_hi":1239,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1127,"Elo_lo":1041,"Elo_hi":1209,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1122,"Elo_lo":1063,"Elo_hi":1185,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1115,"Elo_lo":1061,"Elo_hi":1168,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1106,"Elo_lo":1051,"Elo_hi":1165,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1090,"Elo_lo":1030,"Elo_hi":1144,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1087,"Elo_lo":1018,"Elo_hi":1158,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1077,"Elo_lo":1027,"Elo_hi":1126,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1076,"Elo_lo":1016,"Elo_hi":1136,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1068,"Elo_lo":991,"Elo_hi":1147,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1066,"Elo_lo":1020,"Elo_hi":1109,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1064,"Elo_lo":1009,"Elo_hi":1116,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1060,"Elo_lo":997,"Elo_hi":1120,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1055,"Elo_lo":1011,"Elo_hi":1097,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1041,"Elo_lo":1016,"Elo_hi":1068,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1033,"Elo_lo":992,"Elo_hi":1072,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1029,"Elo_lo":965,"Elo_hi":1094,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1020,"Elo_lo":945,"Elo_hi":1092,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1018,"Elo_lo":993,"Elo_hi":1046,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1010,"Elo_lo":936,"Elo_hi":1079,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":1001,"Elo_lo":953,"Elo_hi":1047,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":990,"Elo_lo":920,"Elo_hi":1057,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT","Elo":984,"Elo_lo":962,"Elo_hi":1008,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":920,"Elo_lo":826,"Elo_hi":1001,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":890,"Elo_lo":848,"Elo_hi":934,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":799,"Elo_lo":730,"Elo_hi":864,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":619,"Elo_lo":489,"Elo_hi":730,"n_targets":46,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":487,"Elo_lo":421,"Elo_hi":546,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1126,"Elo_lo":1064,"Elo_hi":1189,"n_targets":45,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1110,"Elo_lo":1045,"Elo_hi":1180,"n_targets":45,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1080,"Elo_lo":1034,"Elo_hi":1130,"n_targets":45,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1053,"Elo_lo":1002,"Elo_hi":1104,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1053,"Elo_lo":1008,"Elo_hi":1099,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1040,"Elo_lo":992,"Elo_hi":1088,"n_targets":45,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1040,"Elo_lo":995,"Elo_hi":1085,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1034,"Elo_lo":980,"Elo_hi":1084,"n_targets":45,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1028,"Elo_lo":971,"Elo_hi":1083,"n_targets":45,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":45,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":999,"Elo_lo":934,"Elo_hi":1064,"n_targets":44,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":994,"Elo_lo":922,"Elo_hi":1067,"n_targets":45,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT","Elo":991,"Elo_lo":972,"Elo_hi":1008,"n_targets":45,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":986,"Elo_lo":928,"Elo_hi":1047,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":985,"Elo_lo":933,"Elo_hi":1035,"n_targets":45,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":977,"Elo_lo":917,"Elo_hi":1033,"n_targets":45,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":976,"Elo_lo":913,"Elo_hi":1037,"n_targets":45,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":970,"Elo_lo":935,"Elo_hi":1000,"n_targets":45,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":953,"Elo_lo":890,"Elo_hi":1016,"n_targets":45,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":949,"Elo_lo":908,"Elo_hi":990,"n_targets":45,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":947,"Elo_lo":886,"Elo_hi":1012,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":934,"Elo_lo":883,"Elo_hi":983,"n_targets":45,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":929,"Elo_lo":885,"Elo_hi":973,"n_targets":45,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":921,"Elo_lo":864,"Elo_hi":976,"n_targets":45,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":911,"Elo_lo":863,"Elo_hi":952,"n_targets":45,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":899,"Elo_lo":850,"Elo_hi":945,"n_targets":45,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":889,"Elo_lo":845,"Elo_hi":932,"n_targets":45,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":881,"Elo_lo":819,"Elo_hi":938,"n_targets":45,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":878,"Elo_lo":828,"Elo_hi":931,"n_targets":45,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":818,"Elo_lo":773,"Elo_hi":861,"n_targets":45,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":797,"Elo_lo":710,"Elo_hi":880,"n_targets":45,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":791,"Elo_lo":738,"Elo_hi":840,"n_targets":45,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":729,"Elo_lo":658,"Elo_hi":796,"n_targets":45,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":726,"Elo_lo":655,"Elo_hi":796,"n_targets":45,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1223,"Elo_lo":1159,"Elo_hi":1286,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1208,"Elo_lo":1157,"Elo_hi":1264,"n_targets":46,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1201,"Elo_lo":1143,"Elo_hi":1258,"n_targets":46,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1140,"Elo_lo":1097,"Elo_hi":1183,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1138,"Elo_lo":1091,"Elo_hi":1180,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1135,"Elo_lo":1084,"Elo_hi":1190,"n_targets":46,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1082,"Elo_lo":1034,"Elo_hi":1128,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1072,"Elo_lo":1022,"Elo_hi":1114,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1071,"Elo_lo":1024,"Elo_hi":1119,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1061,"Elo_lo":1009,"Elo_hi":1113,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1041,"Elo_lo":970,"Elo_hi":1113,"n_targets":46,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1026,"Elo_lo":956,"Elo_hi":1100,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1024,"Elo_lo":951,"Elo_hi":1099,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1020,"Elo_lo":968,"Elo_hi":1070,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1018,"Elo_lo":968,"Elo_hi":1068,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":996,"Elo_lo":936,"Elo_hi":1059,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":995,"Elo_lo":941,"Elo_hi":1054,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":994,"Elo_lo":936,"Elo_hi":1048,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT","Elo":994,"Elo_lo":974,"Elo_hi":1013,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":987,"Elo_lo":935,"Elo_hi":1041,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":983,"Elo_lo":944,"Elo_hi":1024,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":971,"Elo_lo":933,"Elo_hi":1008,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":963,"Elo_lo":932,"Elo_hi":997,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":958,"Elo_lo":925,"Elo_hi":996,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":951,"Elo_lo":905,"Elo_hi":998,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":948,"Elo_lo":897,"Elo_hi":1000,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":946,"Elo_lo":894,"Elo_hi":996,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":939,"Elo_lo":892,"Elo_hi":990,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":852,"Elo_lo":807,"Elo_hi":894,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":844,"Elo_lo":801,"Elo_hi":889,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":841,"Elo_lo":793,"Elo_hi":884,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":801,"Elo_lo":723,"Elo_hi":877,"n_targets":46,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":650,"Elo_lo":589,"Elo_hi":711,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1235,"Elo_lo":1171,"Elo_hi":1307,"n_targets":44,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1218,"Elo_lo":1152,"Elo_hi":1290,"n_targets":44,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1206,"Elo_lo":1138,"Elo_hi":1280,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1192,"Elo_lo":1151,"Elo_hi":1234,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1191,"Elo_lo":1132,"Elo_hi":1255,"n_targets":44,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1173,"Elo_lo":1120,"Elo_hi":1229,"n_targets":44,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1099,"Elo_lo":1027,"Elo_hi":1172,"n_targets":44,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1082,"Elo_lo":1036,"Elo_hi":1127,"n_targets":44,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1075,"Elo_lo":1013,"Elo_hi":1135,"n_targets":43,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1059,"Elo_lo":1001,"Elo_hi":1117,"n_targets":44,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1057,"Elo_lo":1002,"Elo_hi":1111,"n_targets":44,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1052,"Elo_lo":982,"Elo_hi":1119,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1049,"Elo_lo":995,"Elo_hi":1103,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1038,"Elo_lo":984,"Elo_hi":1089,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1034,"Elo_lo":962,"Elo_hi":1104,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1033,"Elo_lo":989,"Elo_hi":1075,"n_targets":44,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1031,"Elo_lo":950,"Elo_hi":1107,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1030,"Elo_lo":968,"Elo_hi":1088,"n_targets":44,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1029,"Elo_lo":945,"Elo_hi":1115,"n_targets":44,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1028,"Elo_lo":961,"Elo_hi":1089,"n_targets":44,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1015,"Elo_lo":970,"Elo_hi":1062,"n_targets":44,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1006,"Elo_lo":977,"Elo_hi":1037,"n_targets":44,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":996,"Elo_lo":933,"Elo_hi":1060,"n_targets":44,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":995,"Elo_lo":931,"Elo_hi":1052,"n_targets":44,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT","Elo":988,"Elo_lo":972,"Elo_hi":1003,"n_targets":44,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":987,"Elo_lo":954,"Elo_hi":1020,"n_targets":44,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":930,"Elo_lo":878,"Elo_hi":976,"n_targets":44,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":926,"Elo_lo":871,"Elo_hi":980,"n_targets":44,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":917,"Elo_lo":876,"Elo_hi":957,"n_targets":44,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":889,"Elo_lo":843,"Elo_hi":929,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":888,"Elo_lo":833,"Elo_hi":943,"n_targets":44,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":798,"Elo_lo":704,"Elo_hi":882,"n_targets":44,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":589,"Elo_lo":529,"Elo_hi":642,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1257,"Elo_lo":1198,"Elo_hi":1323,"n_targets":36,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1202,"Elo_lo":1135,"Elo_hi":1265,"n_targets":36,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1199,"Elo_lo":1140,"Elo_hi":1260,"n_targets":36,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1194,"Elo_lo":1144,"Elo_hi":1249,"n_targets":36,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1169,"Elo_lo":1116,"Elo_hi":1224,"n_targets":36,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1157,"Elo_lo":1094,"Elo_hi":1225,"n_targets":24,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1117,"Elo_lo":1042,"Elo_hi":1203,"n_targets":36,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1110,"Elo_lo":1056,"Elo_hi":1169,"n_targets":36,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1098,"Elo_lo":1046,"Elo_hi":1147,"n_targets":36,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1077,"Elo_lo":998,"Elo_hi":1155,"n_targets":36,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1074,"Elo_lo":990,"Elo_hi":1154,"n_targets":24,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1063,"Elo_lo":1005,"Elo_hi":1121,"n_targets":36,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1056,"Elo_lo":999,"Elo_hi":1119,"n_targets":36,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1055,"Elo_lo":982,"Elo_hi":1127,"n_targets":36,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1055,"Elo_lo":1000,"Elo_hi":1109,"n_targets":36,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1043,"Elo_lo":950,"Elo_hi":1130,"n_targets":36,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1043,"Elo_lo":956,"Elo_hi":1126,"n_targets":24,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1037,"Elo_lo":999,"Elo_hi":1079,"n_targets":36,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1035,"Elo_lo":960,"Elo_hi":1107,"n_targets":36,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1024,"Elo_lo":966,"Elo_hi":1077,"n_targets":36,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1023,"Elo_lo":960,"Elo_hi":1081,"n_targets":36,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1019,"Elo_lo":984,"Elo_hi":1054,"n_targets":36,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1008,"Elo_lo":972,"Elo_hi":1045,"n_targets":36,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1001,"Elo_lo":947,"Elo_hi":1062,"n_targets":36,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":36,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":996,"Elo_lo":943,"Elo_hi":1048,"n_targets":36,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT","Elo":991,"Elo_lo":971,"Elo_hi":1008,"n_targets":36,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":984,"Elo_lo":940,"Elo_hi":1031,"n_targets":36,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":970,"Elo_lo":885,"Elo_hi":1053,"n_targets":36,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":934,"Elo_lo":879,"Elo_hi":994,"n_targets":36,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":923,"Elo_lo":851,"Elo_hi":989,"n_targets":36,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":870,"Elo_lo":759,"Elo_hi":970,"n_targets":36,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":857,"Elo_lo":792,"Elo_hi":918,"n_targets":36,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":543,"Elo_lo":484,"Elo_hi":600,"n_targets":36,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1206,"Elo_lo":1110,"Elo_hi":1318,"n_targets":22,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1170,"Elo_lo":1098,"Elo_hi":1259,"n_targets":22,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1136,"Elo_lo":1037,"Elo_hi":1237,"n_targets":22,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1133,"Elo_lo":1011,"Elo_hi":1258,"n_targets":14,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1120,"Elo_lo":1033,"Elo_hi":1206,"n_targets":22,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1103,"Elo_lo":1001,"Elo_hi":1201,"n_targets":22,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1076,"Elo_lo":1000,"Elo_hi":1163,"n_targets":22,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1071,"Elo_lo":965,"Elo_hi":1174,"n_targets":14,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1065,"Elo_lo":990,"Elo_hi":1142,"n_targets":22,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1050,"Elo_lo":957,"Elo_hi":1140,"n_targets":22,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1046,"Elo_lo":969,"Elo_hi":1123,"n_targets":22,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1034,"Elo_lo":933,"Elo_hi":1145,"n_targets":22,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1029,"Elo_lo":929,"Elo_hi":1137,"n_targets":22,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1027,"Elo_lo":996,"Elo_hi":1060,"n_targets":22,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1022,"Elo_lo":922,"Elo_hi":1121,"n_targets":22,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1020,"Elo_lo":928,"Elo_hi":1105,"n_targets":22,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1016,"Elo_lo":958,"Elo_hi":1070,"n_targets":22,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1005,"Elo_lo":866,"Elo_hi":1136,"n_targets":22,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":22,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT","Elo":999,"Elo_lo":974,"Elo_hi":1022,"n_targets":22,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":995,"Elo_lo":918,"Elo_hi":1067,"n_targets":22,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":995,"Elo_lo":904,"Elo_hi":1086,"n_targets":22,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":991,"Elo_lo":901,"Elo_hi":1075,"n_targets":22,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":985,"Elo_lo":915,"Elo_hi":1054,"n_targets":22,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":974,"Elo_lo":840,"Elo_hi":1110,"n_targets":14,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":970,"Elo_lo":917,"Elo_hi":1025,"n_targets":22,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":939,"Elo_lo":862,"Elo_hi":1012,"n_targets":22,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR","Elo":937,"Elo_lo":798,"Elo_hi":1055,"n_targets":22,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":933,"Elo_lo":836,"Elo_hi":1023,"n_targets":22,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":851,"Elo_lo":759,"Elo_hi":938,"n_targets":22,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":814,"Elo_lo":688,"Elo_hi":945,"n_targets":22,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":802,"Elo_lo":650,"Elo_hi":935,"n_targets":22,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":756,"Elo_lo":652,"Elo_hi":847,"n_targets":22,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":472,"Elo_lo":372,"Elo_hi":564,"n_targets":22,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1206,"Elo_lo":1078,"Elo_hi":1352,"n_targets":17,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1205,"Elo_lo":1126,"Elo_hi":1308,"n_targets":17,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1158,"Elo_lo":1049,"Elo_hi":1274,"n_targets":17,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1115,"Elo_lo":992,"Elo_hi":1238,"n_targets":10,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1112,"Elo_lo":972,"Elo_hi":1269,"n_targets":17,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1111,"Elo_lo":986,"Elo_hi":1245,"n_targets":17,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1091,"Elo_lo":989,"Elo_hi":1216,"n_targets":17,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1083,"Elo_lo":970,"Elo_hi":1216,"n_targets":17,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1082,"Elo_lo":957,"Elo_hi":1222,"n_targets":17,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1062,"Elo_lo":903,"Elo_hi":1231,"n_targets":10,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1051,"Elo_lo":943,"Elo_hi":1162,"n_targets":17,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1015,"Elo_lo":901,"Elo_hi":1126,"n_targets":17,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1001,"Elo_lo":914,"Elo_hi":1087,"n_targets":17,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":17,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT","Elo":997,"Elo_lo":963,"Elo_hi":1035,"n_targets":17,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":993,"Elo_lo":952,"Elo_hi":1031,"n_targets":17,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":985,"Elo_lo":844,"Elo_hi":1129,"n_targets":17,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":984,"Elo_lo":898,"Elo_hi":1057,"n_targets":17,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":980,"Elo_lo":855,"Elo_hi":1113,"n_targets":17,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":979,"Elo_lo":909,"Elo_hi":1046,"n_targets":17,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":978,"Elo_lo":929,"Elo_hi":1025,"n_targets":17,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":970,"Elo_lo":826,"Elo_hi":1110,"n_targets":17,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":958,"Elo_lo":834,"Elo_hi":1083,"n_targets":17,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":930,"Elo_lo":732,"Elo_hi":1122,"n_targets":17,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":926,"Elo_lo":849,"Elo_hi":998,"n_targets":17,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":912,"Elo_lo":711,"Elo_hi":1114,"n_targets":10,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":863,"Elo_lo":742,"Elo_hi":993,"n_targets":17,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":856,"Elo_lo":726,"Elo_hi":1014,"n_targets":17,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":838,"Elo_lo":688,"Elo_hi":966,"n_targets":17,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR","Elo":794,"Elo_lo":553,"Elo_hi":986,"n_targets":17,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":713,"Elo_lo":593,"Elo_hi":817,"n_targets":17,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":700,"Elo_lo":580,"Elo_hi":813,"n_targets":17,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":537,"Elo_lo":288,"Elo_hi":758,"n_targets":17,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":391,"Elo_lo":241,"Elo_hi":523,"n_targets":17,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1196,"Elo_lo":1150,"Elo_hi":1245,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1185,"Elo_lo":1117,"Elo_hi":1251,"n_targets":46,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1177,"Elo_lo":1114,"Elo_hi":1242,"n_targets":46,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1170,"Elo_lo":1112,"Elo_hi":1228,"n_targets":46,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1153,"Elo_lo":1094,"Elo_hi":1213,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1122,"Elo_lo":1065,"Elo_hi":1181,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1107,"Elo_lo":1016,"Elo_hi":1197,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1095,"Elo_lo":1044,"Elo_hi":1145,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1089,"Elo_lo":1032,"Elo_hi":1150,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1075,"Elo_lo":1018,"Elo_hi":1132,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1066,"Elo_lo":1022,"Elo_hi":1111,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1065,"Elo_lo":1010,"Elo_hi":1115,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1061,"Elo_lo":994,"Elo_hi":1127,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1060,"Elo_lo":1021,"Elo_hi":1098,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1046,"Elo_lo":995,"Elo_hi":1095,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1045,"Elo_lo":1000,"Elo_hi":1090,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1041,"Elo_lo":982,"Elo_hi":1098,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1035,"Elo_lo":996,"Elo_hi":1076,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1032,"Elo_lo":957,"Elo_hi":1106,"n_targets":46,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1032,"Elo_lo":999,"Elo_hi":1064,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1015,"Elo_lo":992,"Elo_hi":1041,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1008,"Elo_lo":938,"Elo_hi":1077,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1006,"Elo_lo":933,"Elo_hi":1078,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":995,"Elo_lo":931,"Elo_hi":1059,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT","Elo":987,"Elo_lo":966,"Elo_hi":1007,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":983,"Elo_lo":942,"Elo_hi":1024,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":977,"Elo_lo":911,"Elo_hi":1040,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":977,"Elo_lo":899,"Elo_hi":1055,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":880,"Elo_lo":836,"Elo_hi":924,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":811,"Elo_lo":712,"Elo_hi":897,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":738,"Elo_lo":673,"Elo_hi":801,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":631,"Elo_lo":525,"Elo_hi":731,"n_targets":46,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":523,"Elo_lo":469,"Elo_hi":569,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1128,"Elo_lo":1066,"Elo_hi":1188,"n_targets":43,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1107,"Elo_lo":1041,"Elo_hi":1172,"n_targets":43,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1086,"Elo_lo":1041,"Elo_hi":1135,"n_targets":43,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1044,"Elo_lo":996,"Elo_hi":1092,"n_targets":43,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1038,"Elo_lo":985,"Elo_hi":1084,"n_targets":43,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1028,"Elo_lo":968,"Elo_hi":1085,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1014,"Elo_lo":959,"Elo_hi":1065,"n_targets":43,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1012,"Elo_lo":950,"Elo_hi":1076,"n_targets":43,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":43,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":997,"Elo_lo":929,"Elo_hi":1059,"n_targets":42,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"XT","Elo":994,"Elo_lo":973,"Elo_hi":1015,"n_targets":43,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":992,"Elo_lo":927,"Elo_hi":1053,"n_targets":43,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":991,"Elo_lo":916,"Elo_hi":1068,"n_targets":43,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":984,"Elo_lo":927,"Elo_hi":1047,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":983,"Elo_lo":923,"Elo_hi":1045,"n_targets":43,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":977,"Elo_lo":922,"Elo_hi":1036,"n_targets":43,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":965,"Elo_lo":911,"Elo_hi":1019,"n_targets":43,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":962,"Elo_lo":925,"Elo_hi":999,"n_targets":43,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":952,"Elo_lo":890,"Elo_hi":1009,"n_targets":43,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":948,"Elo_lo":899,"Elo_hi":996,"n_targets":43,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":942,"Elo_lo":877,"Elo_hi":1004,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":933,"Elo_lo":885,"Elo_hi":980,"n_targets":43,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":927,"Elo_lo":871,"Elo_hi":975,"n_targets":43,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":923,"Elo_lo":865,"Elo_hi":978,"n_targets":43,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":912,"Elo_lo":862,"Elo_hi":956,"n_targets":43,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":905,"Elo_lo":854,"Elo_hi":951,"n_targets":43,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":899,"Elo_lo":856,"Elo_hi":939,"n_targets":43,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":884,"Elo_lo":832,"Elo_hi":936,"n_targets":43,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":883,"Elo_lo":819,"Elo_hi":945,"n_targets":43,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":836,"Elo_lo":789,"Elo_hi":881,"n_targets":43,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":800,"Elo_lo":711,"Elo_hi":886,"n_targets":43,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":796,"Elo_lo":743,"Elo_hi":848,"n_targets":43,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":734,"Elo_lo":667,"Elo_hi":801,"n_targets":43,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":729,"Elo_lo":661,"Elo_hi":794,"n_targets":43,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1202,"Elo_lo":1145,"Elo_hi":1255,"n_targets":44,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1198,"Elo_lo":1138,"Elo_hi":1255,"n_targets":44,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1183,"Elo_lo":1112,"Elo_hi":1253,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1133,"Elo_lo":1090,"Elo_hi":1176,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1131,"Elo_lo":1077,"Elo_hi":1182,"n_targets":44,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1082,"Elo_lo":1030,"Elo_hi":1136,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1079,"Elo_lo":1020,"Elo_hi":1132,"n_targets":44,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1058,"Elo_lo":1010,"Elo_hi":1102,"n_targets":44,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1056,"Elo_lo":1002,"Elo_hi":1104,"n_targets":43,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1046,"Elo_lo":971,"Elo_hi":1125,"n_targets":44,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1024,"Elo_lo":949,"Elo_hi":1091,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1022,"Elo_lo":950,"Elo_hi":1100,"n_targets":44,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1012,"Elo_lo":955,"Elo_hi":1070,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1004,"Elo_lo":942,"Elo_hi":1064,"n_targets":44,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":998,"Elo_lo":944,"Elo_hi":1050,"n_targets":44,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"XT","Elo":996,"Elo_lo":973,"Elo_hi":1017,"n_targets":44,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":995,"Elo_lo":934,"Elo_hi":1054,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":982,"Elo_lo":922,"Elo_hi":1045,"n_targets":44,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":978,"Elo_lo":939,"Elo_hi":1018,"n_targets":44,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":978,"Elo_lo":940,"Elo_hi":1014,"n_targets":44,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":968,"Elo_lo":937,"Elo_hi":1000,"n_targets":44,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":968,"Elo_lo":907,"Elo_hi":1024,"n_targets":44,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":967,"Elo_lo":909,"Elo_hi":1027,"n_targets":44,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":958,"Elo_lo":910,"Elo_hi":1008,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":952,"Elo_lo":898,"Elo_hi":1004,"n_targets":44,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":952,"Elo_lo":913,"Elo_hi":987,"n_targets":44,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":933,"Elo_lo":883,"Elo_hi":988,"n_targets":44,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":931,"Elo_lo":880,"Elo_hi":980,"n_targets":44,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":862,"Elo_lo":816,"Elo_hi":907,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":855,"Elo_lo":806,"Elo_hi":901,"n_targets":44,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":845,"Elo_lo":808,"Elo_hi":884,"n_targets":44,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":822,"Elo_lo":743,"Elo_hi":900,"n_targets":44,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":675,"Elo_lo":614,"Elo_hi":732,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1196,"Elo_lo":1121,"Elo_hi":1266,"n_targets":42,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1182,"Elo_lo":1107,"Elo_hi":1253,"n_targets":42,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1179,"Elo_lo":1136,"Elo_hi":1221,"n_targets":42,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1153,"Elo_lo":1078,"Elo_hi":1232,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1147,"Elo_lo":1084,"Elo_hi":1214,"n_targets":42,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1067,"Elo_lo":992,"Elo_hi":1137,"n_targets":42,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1058,"Elo_lo":998,"Elo_hi":1118,"n_targets":41,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1044,"Elo_lo":986,"Elo_hi":1097,"n_targets":42,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1041,"Elo_lo":985,"Elo_hi":1095,"n_targets":42,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1037,"Elo_lo":973,"Elo_hi":1098,"n_targets":42,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1035,"Elo_lo":943,"Elo_hi":1122,"n_targets":42,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1020,"Elo_lo":964,"Elo_hi":1078,"n_targets":42,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1020,"Elo_lo":942,"Elo_hi":1095,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1019,"Elo_lo":946,"Elo_hi":1085,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1018,"Elo_lo":965,"Elo_hi":1071,"n_targets":42,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1012,"Elo_lo":965,"Elo_hi":1053,"n_targets":42,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1008,"Elo_lo":927,"Elo_hi":1090,"n_targets":42,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":42,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":998,"Elo_lo":936,"Elo_hi":1059,"n_targets":42,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":995,"Elo_lo":924,"Elo_hi":1061,"n_targets":42,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":987,"Elo_lo":940,"Elo_hi":1034,"n_targets":42,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"XT","Elo":985,"Elo_lo":969,"Elo_hi":1003,"n_targets":42,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":982,"Elo_lo":949,"Elo_hi":1016,"n_targets":42,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":979,"Elo_lo":944,"Elo_hi":1013,"n_targets":42,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":970,"Elo_lo":900,"Elo_hi":1035,"n_targets":42,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":966,"Elo_lo":910,"Elo_hi":1024,"n_targets":42,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":947,"Elo_lo":873,"Elo_hi":1022,"n_targets":42,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":918,"Elo_lo":872,"Elo_hi":964,"n_targets":42,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":900,"Elo_lo":860,"Elo_hi":937,"n_targets":42,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":881,"Elo_lo":820,"Elo_hi":943,"n_targets":42,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":876,"Elo_lo":832,"Elo_hi":919,"n_targets":42,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":872,"Elo_lo":819,"Elo_hi":924,"n_targets":42,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":805,"Elo_lo":718,"Elo_hi":887,"n_targets":42,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":615,"Elo_lo":562,"Elo_hi":661,"n_targets":42,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1246,"Elo_lo":1185,"Elo_hi":1316,"n_targets":34,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1196,"Elo_lo":1144,"Elo_hi":1249,"n_targets":34,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1190,"Elo_lo":1129,"Elo_hi":1248,"n_targets":34,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1182,"Elo_lo":1117,"Elo_hi":1242,"n_targets":34,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1117,"Elo_lo":1035,"Elo_hi":1195,"n_targets":24,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1105,"Elo_lo":1028,"Elo_hi":1178,"n_targets":34,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1077,"Elo_lo":989,"Elo_hi":1167,"n_targets":34,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1072,"Elo_lo":986,"Elo_hi":1147,"n_targets":34,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1068,"Elo_lo":1010,"Elo_hi":1129,"n_targets":34,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1068,"Elo_lo":986,"Elo_hi":1148,"n_targets":24,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1063,"Elo_lo":971,"Elo_hi":1156,"n_targets":34,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1058,"Elo_lo":997,"Elo_hi":1119,"n_targets":34,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1053,"Elo_lo":988,"Elo_hi":1120,"n_targets":34,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1047,"Elo_lo":989,"Elo_hi":1103,"n_targets":34,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1044,"Elo_lo":968,"Elo_hi":1114,"n_targets":34,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1038,"Elo_lo":958,"Elo_hi":1110,"n_targets":34,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1037,"Elo_lo":950,"Elo_hi":1119,"n_targets":24,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1027,"Elo_lo":953,"Elo_hi":1101,"n_targets":34,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1019,"Elo_lo":949,"Elo_hi":1081,"n_targets":34,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1008,"Elo_lo":939,"Elo_hi":1076,"n_targets":34,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1005,"Elo_lo":945,"Elo_hi":1063,"n_targets":34,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":34,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":999,"Elo_lo":959,"Elo_hi":1040,"n_targets":34,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":996,"Elo_lo":948,"Elo_hi":1044,"n_targets":34,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":993,"Elo_lo":939,"Elo_hi":1047,"n_targets":34,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"XT","Elo":992,"Elo_lo":974,"Elo_hi":1009,"n_targets":34,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":990,"Elo_lo":949,"Elo_hi":1030,"n_targets":34,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":988,"Elo_lo":902,"Elo_hi":1077,"n_targets":34,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":962,"Elo_lo":905,"Elo_hi":1024,"n_targets":34,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":955,"Elo_lo":900,"Elo_hi":1010,"n_targets":34,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":940,"Elo_lo":865,"Elo_hi":1013,"n_targets":34,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":887,"Elo_lo":778,"Elo_hi":981,"n_targets":34,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":863,"Elo_lo":799,"Elo_hi":923,"n_targets":34,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":581,"Elo_lo":513,"Elo_hi":636,"n_targets":34,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1201,"Elo_lo":1103,"Elo_hi":1313,"n_targets":22,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1160,"Elo_lo":1091,"Elo_hi":1248,"n_targets":22,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1133,"Elo_lo":1036,"Elo_hi":1235,"n_targets":22,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1096,"Elo_lo":998,"Elo_hi":1193,"n_targets":22,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1096,"Elo_lo":1000,"Elo_hi":1192,"n_targets":22,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1082,"Elo_lo":954,"Elo_hi":1215,"n_targets":14,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1071,"Elo_lo":967,"Elo_hi":1177,"n_targets":14,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1060,"Elo_lo":985,"Elo_hi":1135,"n_targets":22,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1044,"Elo_lo":966,"Elo_hi":1116,"n_targets":22,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1034,"Elo_lo":939,"Elo_hi":1123,"n_targets":22,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1033,"Elo_lo":947,"Elo_hi":1134,"n_targets":22,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1031,"Elo_lo":931,"Elo_hi":1142,"n_targets":22,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1028,"Elo_lo":933,"Elo_hi":1114,"n_targets":22,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1020,"Elo_lo":921,"Elo_hi":1126,"n_targets":22,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1019,"Elo_lo":923,"Elo_hi":1112,"n_targets":22,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1011,"Elo_lo":961,"Elo_hi":1053,"n_targets":22,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1009,"Elo_lo":943,"Elo_hi":1067,"n_targets":22,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1001,"Elo_lo":863,"Elo_hi":1128,"n_targets":22,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":22,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"XT","Elo":996,"Elo_lo":973,"Elo_hi":1019,"n_targets":22,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":983,"Elo_lo":910,"Elo_hi":1052,"n_targets":22,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":982,"Elo_lo":896,"Elo_hi":1065,"n_targets":22,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":977,"Elo_lo":924,"Elo_hi":1033,"n_targets":22,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":973,"Elo_lo":840,"Elo_hi":1106,"n_targets":14,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":970,"Elo_lo":873,"Elo_hi":1063,"n_targets":22,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":952,"Elo_lo":881,"Elo_hi":1022,"n_targets":22,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":937,"Elo_lo":861,"Elo_hi":1010,"n_targets":22,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"LR","Elo":927,"Elo_lo":793,"Elo_hi":1042,"n_targets":22,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":927,"Elo_lo":831,"Elo_hi":1014,"n_targets":22,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":867,"Elo_lo":782,"Elo_hi":953,"n_targets":22,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":814,"Elo_lo":693,"Elo_hi":945,"n_targets":22,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":802,"Elo_lo":656,"Elo_hi":931,"n_targets":22,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":768,"Elo_lo":664,"Elo_hi":856,"n_targets":22,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":490,"Elo_lo":386,"Elo_hi":582,"n_targets":22,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1206,"Elo_lo":1077,"Elo_hi":1351,"n_targets":17,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1205,"Elo_lo":1126,"Elo_hi":1308,"n_targets":17,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1158,"Elo_lo":1048,"Elo_hi":1273,"n_targets":17,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1117,"Elo_lo":994,"Elo_hi":1241,"n_targets":10,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1111,"Elo_lo":972,"Elo_hi":1268,"n_targets":17,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1110,"Elo_lo":986,"Elo_hi":1243,"n_targets":17,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1083,"Elo_lo":970,"Elo_hi":1217,"n_targets":17,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1082,"Elo_lo":957,"Elo_hi":1222,"n_targets":17,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1076,"Elo_lo":971,"Elo_hi":1206,"n_targets":17,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1064,"Elo_lo":905,"Elo_hi":1232,"n_targets":10,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1051,"Elo_lo":943,"Elo_hi":1162,"n_targets":17,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1015,"Elo_lo":901,"Elo_hi":1127,"n_targets":17,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1001,"Elo_lo":914,"Elo_hi":1088,"n_targets":17,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":17,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT","Elo":997,"Elo_lo":963,"Elo_hi":1035,"n_targets":17,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":993,"Elo_lo":952,"Elo_hi":1031,"n_targets":17,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":985,"Elo_lo":845,"Elo_hi":1130,"n_targets":17,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":984,"Elo_lo":898,"Elo_hi":1057,"n_targets":17,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":980,"Elo_lo":855,"Elo_hi":1112,"n_targets":17,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":979,"Elo_lo":909,"Elo_hi":1046,"n_targets":17,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":978,"Elo_lo":929,"Elo_hi":1025,"n_targets":17,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":970,"Elo_lo":826,"Elo_hi":1110,"n_targets":17,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":958,"Elo_lo":834,"Elo_hi":1084,"n_targets":17,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":930,"Elo_lo":732,"Elo_hi":1123,"n_targets":17,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":926,"Elo_lo":849,"Elo_hi":998,"n_targets":17,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":911,"Elo_lo":711,"Elo_hi":1113,"n_targets":10,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":863,"Elo_lo":742,"Elo_hi":992,"n_targets":17,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":857,"Elo_lo":727,"Elo_hi":1014,"n_targets":17,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":839,"Elo_lo":691,"Elo_hi":967,"n_targets":17,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR","Elo":794,"Elo_lo":554,"Elo_hi":986,"n_targets":17,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":714,"Elo_lo":595,"Elo_hi":818,"n_targets":17,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":703,"Elo_lo":585,"Elo_hi":816,"n_targets":17,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":538,"Elo_lo":291,"Elo_hi":759,"n_targets":17,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":393,"Elo_lo":242,"Elo_hi":523,"n_targets":17,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1182,"Elo_lo":1142,"Elo_hi":1228,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1163,"Elo_lo":1094,"Elo_hi":1228,"n_targets":44,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1146,"Elo_lo":1079,"Elo_hi":1207,"n_targets":44,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1145,"Elo_lo":1089,"Elo_hi":1205,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1145,"Elo_lo":1084,"Elo_hi":1210,"n_targets":44,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"AUTOGLUON","Elo":1144,"Elo_lo":1088,"Elo_hi":1200,"n_targets":44,"id":"AUTOGLUON","display":"AutoGluon","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"AutoGluon automates tabular learning by training multiple models and combining their predictions in ensembles. Here it runs its native extreme preset within a one-hour budget as a separate AutoML reference.","tuning":null,"category":"AutoML","color":"#111827","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1068,"Elo_lo":1013,"Elo_hi":1123,"n_targets":44,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1064,"Elo_lo":1019,"Elo_hi":1115,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1047,"Elo_lo":957,"Elo_hi":1131,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1046,"Elo_lo":994,"Elo_hi":1095,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1041,"Elo_lo":999,"Elo_hi":1086,"n_targets":44,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1040,"Elo_lo":995,"Elo_hi":1086,"n_targets":44,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1039,"Elo_lo":966,"Elo_hi":1117,"n_targets":44,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1034,"Elo_lo":961,"Elo_hi":1102,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1032,"Elo_lo":962,"Elo_hi":1100,"n_targets":44,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1031,"Elo_lo":972,"Elo_hi":1087,"n_targets":44,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1018,"Elo_lo":961,"Elo_hi":1072,"n_targets":44,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1005,"Elo_lo":953,"Elo_hi":1056,"n_targets":44,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1005,"Elo_lo":962,"Elo_hi":1046,"n_targets":44,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"LR","Elo":992,"Elo_lo":922,"Elo_hi":1058,"n_targets":44,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":991,"Elo_lo":956,"Elo_hi":1022,"n_targets":44,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":990,"Elo_lo":950,"Elo_hi":1031,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"XT","Elo":987,"Elo_lo":966,"Elo_hi":1006,"n_targets":44,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":986,"Elo_lo":916,"Elo_hi":1052,"n_targets":44,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":985,"Elo_lo":919,"Elo_hi":1052,"n_targets":44,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":984,"Elo_lo":908,"Elo_hi":1057,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":967,"Elo_lo":919,"Elo_hi":1014,"n_targets":44,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":964,"Elo_lo":897,"Elo_hi":1022,"n_targets":43,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":916,"Elo_lo":835,"Elo_hi":997,"n_targets":44,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":876,"Elo_lo":831,"Elo_hi":920,"n_targets":44,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":804,"Elo_lo":719,"Elo_hi":889,"n_targets":44,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":753,"Elo_lo":691,"Elo_hi":810,"n_targets":44,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":654,"Elo_lo":559,"Elo_hi":745,"n_targets":44,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":557,"Elo_lo":503,"Elo_hi":604,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false}],"analysis_views":{"strict":{"reference":[{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1239,"Elo_lo":1182,"Elo_hi":1302,"n_targets":44,"f1_macro":0.737513,"rmse":1.75962,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1226,"Elo_lo":1159,"Elo_hi":1290,"n_targets":44,"f1_macro":0.740375,"rmse":1.8159,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1210,"Elo_lo":1153,"Elo_hi":1268,"n_targets":32,"f1_macro":0.73223,"rmse":null,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1210,"Elo_lo":1145,"Elo_hi":1280,"n_targets":44,"f1_macro":0.719262,"rmse":1.7985,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1199,"Elo_lo":1136,"Elo_hi":1268,"n_targets":44,"f1_macro":0.713243,"rmse":1.85064,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1193,"Elo_lo":1151,"Elo_hi":1237,"n_targets":44,"f1_macro":0.719811,"rmse":1.80462,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1158,"Elo_lo":1107,"Elo_hi":1206,"n_targets":44,"f1_macro":0.702762,"rmse":1.74289,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1110,"Elo_lo":1032,"Elo_hi":1184,"n_targets":44,"f1_macro":0.672508,"rmse":1.83574,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1072,"Elo_lo":1024,"Elo_hi":1119,"n_targets":44,"f1_macro":0.673781,"rmse":1.83726,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1063,"Elo_lo":998,"Elo_hi":1125,"n_targets":44,"f1_macro":0.707116,"rmse":1.95582,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1060,"Elo_lo":1001,"Elo_hi":1120,"n_targets":43,"f1_macro":0.685468,"rmse":1.89663,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1055,"Elo_lo":1002,"Elo_hi":1113,"n_targets":44,"f1_macro":0.685489,"rmse":1.96288,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1050,"Elo_lo":975,"Elo_hi":1120,"n_targets":44,"f1_macro":0.692084,"rmse":1.92319,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1046,"Elo_lo":993,"Elo_hi":1098,"n_targets":44,"f1_macro":0.654833,"rmse":1.8178,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"AUTOGLUON","Elo":1040,"Elo_lo":992,"Elo_hi":1090,"n_targets":44,"f1_macro":0.684564,"rmse":2.07258,"id":"AUTOGLUON","display":"AutoGluon","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"AutoGluon automates tabular learning by training multiple models and combining their predictions in ensembles. Here it runs its native extreme preset within a one-hour budget as a separate AutoML reference.","tuning":null,"category":"AutoML","color":"#111827","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1040,"Elo_lo":984,"Elo_hi":1095,"n_targets":44,"f1_macro":0.686295,"rmse":2.11554,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1040,"Elo_lo":990,"Elo_hi":1092,"n_targets":44,"f1_macro":0.684965,"rmse":2.09472,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1039,"Elo_lo":988,"Elo_hi":1094,"n_targets":44,"f1_macro":0.683193,"rmse":2.12524,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1035,"Elo_lo":958,"Elo_hi":1103,"n_targets":32,"f1_macro":0.662519,"rmse":null,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1023,"Elo_lo":942,"Elo_hi":1094,"n_targets":32,"f1_macro":0.655745,"rmse":null,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1018,"Elo_lo":954,"Elo_hi":1085,"n_targets":44,"f1_macro":0.629495,"rmse":1.79114,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1014,"Elo_lo":972,"Elo_hi":1051,"n_targets":44,"f1_macro":0.662073,"rmse":2.02183,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1011,"Elo_lo":969,"Elo_hi":1061,"n_targets":44,"f1_macro":0.668661,"rmse":2.01502,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1008,"Elo_lo":940,"Elo_hi":1071,"n_targets":44,"f1_macro":0.689285,"rmse":2.17055,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"f1_macro":0.65413,"rmse":1.93045,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":991,"Elo_lo":962,"Elo_hi":1023,"n_targets":44,"f1_macro":0.647685,"rmse":1.99831,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":980,"Elo_lo":946,"Elo_hi":1017,"n_targets":44,"f1_macro":0.63455,"rmse":1.97764,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT","Elo":976,"Elo_lo":952,"Elo_hi":998,"n_targets":44,"f1_macro":0.633911,"rmse":1.99815,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":967,"Elo_lo":918,"Elo_hi":1016,"n_targets":44,"f1_macro":0.636697,"rmse":1.90252,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":915,"Elo_lo":860,"Elo_hi":966,"n_targets":44,"f1_macro":0.628504,"rmse":1.88066,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":909,"Elo_lo":863,"Elo_hi":954,"n_targets":44,"f1_macro":0.613568,"rmse":1.92813,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":880,"Elo_lo":832,"Elo_hi":927,"n_targets":44,"f1_macro":0.625905,"rmse":2.10961,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":873,"Elo_lo":821,"Elo_hi":924,"n_targets":44,"f1_macro":0.619756,"rmse":2.18334,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":788,"Elo_lo":687,"Elo_hi":876,"n_targets":44,"f1_macro":0.371345,"rmse":2.16769,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":558,"Elo_lo":487,"Elo_hi":616,"n_targets":44,"f1_macro":0.33199,"rmse":2.23666,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false}],"cell_elo":[{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1153,"Elo_lo":1102,"Elo_hi":1208,"n_targets":45,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1144,"Elo_lo":1087,"Elo_hi":1207,"n_targets":45,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1120,"Elo_lo":1062,"Elo_hi":1183,"n_targets":45,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1116,"Elo_lo":1057,"Elo_hi":1174,"n_targets":45,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1108,"Elo_lo":1061,"Elo_hi":1153,"n_targets":45,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1075,"Elo_lo":1032,"Elo_hi":1122,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1068,"Elo_lo":1021,"Elo_hi":1114,"n_targets":45,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1067,"Elo_lo":1022,"Elo_hi":1109,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1041,"Elo_lo":987,"Elo_hi":1091,"n_targets":45,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1039,"Elo_lo":995,"Elo_hi":1082,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1024,"Elo_lo":976,"Elo_hi":1069,"n_targets":45,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1021,"Elo_lo":980,"Elo_hi":1063,"n_targets":45,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1005,"Elo_lo":942,"Elo_hi":1065,"n_targets":44,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1002,"Elo_lo":940,"Elo_hi":1066,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT","Elo":1001,"Elo_lo":975,"Elo_hi":1024,"n_targets":45,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":45,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":995,"Elo_lo":957,"Elo_hi":1025,"n_targets":45,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":991,"Elo_lo":926,"Elo_hi":1060,"n_targets":45,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":968,"Elo_lo":905,"Elo_hi":1030,"n_targets":45,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":963,"Elo_lo":897,"Elo_hi":1030,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":951,"Elo_lo":897,"Elo_hi":1005,"n_targets":45,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":946,"Elo_lo":898,"Elo_hi":994,"n_targets":45,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":936,"Elo_lo":882,"Elo_hi":988,"n_targets":45,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":928,"Elo_lo":881,"Elo_hi":969,"n_targets":45,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":920,"Elo_lo":867,"Elo_hi":963,"n_targets":45,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":910,"Elo_lo":832,"Elo_hi":985,"n_targets":45,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":902,"Elo_lo":842,"Elo_hi":962,"n_targets":45,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":900,"Elo_lo":840,"Elo_hi":958,"n_targets":45,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":891,"Elo_lo":843,"Elo_hi":933,"n_targets":45,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":890,"Elo_lo":838,"Elo_hi":943,"n_targets":45,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":845,"Elo_lo":799,"Elo_hi":887,"n_targets":45,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":805,"Elo_lo":754,"Elo_hi":853,"n_targets":45,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":740,"Elo_lo":667,"Elo_hi":807,"n_targets":45,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":739,"Elo_lo":666,"Elo_hi":808,"n_targets":45,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1239,"Elo_lo":1177,"Elo_hi":1303,"n_targets":46,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1232,"Elo_lo":1179,"Elo_hi":1289,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1228,"Elo_lo":1175,"Elo_hi":1278,"n_targets":46,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1222,"Elo_lo":1161,"Elo_hi":1283,"n_targets":46,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1212,"Elo_lo":1154,"Elo_hi":1270,"n_targets":46,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1143,"Elo_lo":1097,"Elo_hi":1191,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1122,"Elo_lo":1073,"Elo_hi":1170,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1103,"Elo_lo":1060,"Elo_hi":1151,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1090,"Elo_lo":1014,"Elo_hi":1164,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1086,"Elo_lo":1038,"Elo_hi":1133,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1073,"Elo_lo":1024,"Elo_hi":1120,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1069,"Elo_lo":1022,"Elo_hi":1118,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1061,"Elo_lo":1010,"Elo_hi":1108,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1027,"Elo_lo":980,"Elo_hi":1075,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1023,"Elo_lo":972,"Elo_hi":1074,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1021,"Elo_lo":974,"Elo_hi":1066,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1020,"Elo_lo":946,"Elo_hi":1093,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1009,"Elo_lo":943,"Elo_hi":1080,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1001,"Elo_lo":956,"Elo_hi":1044,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":999,"Elo_lo":944,"Elo_hi":1056,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT","Elo":996,"Elo_lo":972,"Elo_hi":1022,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":984,"Elo_lo":938,"Elo_hi":1028,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":983,"Elo_lo":955,"Elo_hi":1010,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":980,"Elo_lo":930,"Elo_hi":1031,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":979,"Elo_lo":915,"Elo_hi":1038,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":974,"Elo_lo":940,"Elo_hi":1007,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":960,"Elo_lo":929,"Elo_hi":992,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":928,"Elo_lo":879,"Elo_hi":977,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":869,"Elo_lo":793,"Elo_hi":944,"n_targets":46,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":845,"Elo_lo":797,"Elo_hi":891,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":838,"Elo_lo":802,"Elo_hi":872,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":790,"Elo_lo":733,"Elo_hi":842,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":651,"Elo_lo":578,"Elo_hi":719,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1276,"Elo_lo":1212,"Elo_hi":1347,"n_targets":44,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1256,"Elo_lo":1194,"Elo_hi":1321,"n_targets":44,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1255,"Elo_lo":1184,"Elo_hi":1328,"n_targets":44,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1245,"Elo_lo":1169,"Elo_hi":1320,"n_targets":44,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1229,"Elo_lo":1160,"Elo_hi":1300,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1215,"Elo_lo":1171,"Elo_hi":1265,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1153,"Elo_lo":1096,"Elo_hi":1213,"n_targets":44,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1143,"Elo_lo":1069,"Elo_hi":1213,"n_targets":44,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1106,"Elo_lo":1053,"Elo_hi":1162,"n_targets":44,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1105,"Elo_lo":1035,"Elo_hi":1175,"n_targets":44,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1100,"Elo_lo":1040,"Elo_hi":1161,"n_targets":44,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1098,"Elo_lo":1043,"Elo_hi":1156,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1083,"Elo_lo":1019,"Elo_hi":1145,"n_targets":44,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1078,"Elo_lo":1000,"Elo_hi":1151,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1078,"Elo_lo":1019,"Elo_hi":1140,"n_targets":44,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1074,"Elo_lo":1016,"Elo_hi":1131,"n_targets":43,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1074,"Elo_lo":1022,"Elo_hi":1130,"n_targets":44,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1055,"Elo_lo":972,"Elo_hi":1134,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1044,"Elo_lo":993,"Elo_hi":1094,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1044,"Elo_lo":1000,"Elo_hi":1092,"n_targets":44,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1037,"Elo_lo":980,"Elo_hi":1095,"n_targets":44,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1034,"Elo_lo":955,"Elo_hi":1106,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1020,"Elo_lo":984,"Elo_hi":1058,"n_targets":44,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1012,"Elo_lo":940,"Elo_hi":1076,"n_targets":44,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1004,"Elo_lo":979,"Elo_hi":1027,"n_targets":44,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT","Elo":988,"Elo_lo":968,"Elo_hi":1006,"n_targets":44,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":977,"Elo_lo":947,"Elo_hi":1010,"n_targets":44,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":917,"Elo_lo":862,"Elo_hi":967,"n_targets":44,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":908,"Elo_lo":864,"Elo_hi":953,"n_targets":44,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":887,"Elo_lo":836,"Elo_hi":938,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":873,"Elo_lo":810,"Elo_hi":937,"n_targets":44,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":862,"Elo_lo":774,"Elo_hi":943,"n_targets":44,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":581,"Elo_lo":504,"Elo_hi":644,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1283,"Elo_lo":1210,"Elo_hi":1359,"n_targets":36,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1228,"Elo_lo":1156,"Elo_hi":1309,"n_targets":36,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1215,"Elo_lo":1137,"Elo_hi":1297,"n_targets":36,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1213,"Elo_lo":1136,"Elo_hi":1292,"n_targets":36,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1212,"Elo_lo":1153,"Elo_hi":1278,"n_targets":36,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1195,"Elo_lo":1128,"Elo_hi":1265,"n_targets":36,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1179,"Elo_lo":1106,"Elo_hi":1259,"n_targets":24,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1177,"Elo_lo":1113,"Elo_hi":1246,"n_targets":36,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1131,"Elo_lo":1048,"Elo_hi":1220,"n_targets":36,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1127,"Elo_lo":1062,"Elo_hi":1195,"n_targets":36,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1111,"Elo_lo":1042,"Elo_hi":1188,"n_targets":36,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1108,"Elo_lo":1022,"Elo_hi":1198,"n_targets":36,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1108,"Elo_lo":1033,"Elo_hi":1185,"n_targets":36,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1095,"Elo_lo":1026,"Elo_hi":1164,"n_targets":36,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1088,"Elo_lo":1000,"Elo_hi":1176,"n_targets":24,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1081,"Elo_lo":995,"Elo_hi":1160,"n_targets":24,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1077,"Elo_lo":1018,"Elo_hi":1136,"n_targets":36,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1071,"Elo_lo":1007,"Elo_hi":1141,"n_targets":36,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1064,"Elo_lo":994,"Elo_hi":1132,"n_targets":36,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1059,"Elo_lo":987,"Elo_hi":1127,"n_targets":36,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1041,"Elo_lo":953,"Elo_hi":1124,"n_targets":36,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1037,"Elo_lo":996,"Elo_hi":1078,"n_targets":36,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1019,"Elo_lo":953,"Elo_hi":1086,"n_targets":36,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1014,"Elo_lo":956,"Elo_hi":1082,"n_targets":36,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1004,"Elo_lo":968,"Elo_hi":1039,"n_targets":36,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":36,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT","Elo":989,"Elo_lo":970,"Elo_hi":1007,"n_targets":36,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":988,"Elo_lo":940,"Elo_hi":1038,"n_targets":36,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":959,"Elo_lo":904,"Elo_hi":1014,"n_targets":36,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":949,"Elo_lo":908,"Elo_hi":992,"n_targets":36,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":921,"Elo_lo":810,"Elo_hi":1020,"n_targets":36,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":916,"Elo_lo":861,"Elo_hi":973,"n_targets":36,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":841,"Elo_lo":768,"Elo_hi":911,"n_targets":36,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":496,"Elo_lo":422,"Elo_hi":569,"n_targets":36,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1231,"Elo_lo":1132,"Elo_hi":1348,"n_targets":22,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1211,"Elo_lo":1124,"Elo_hi":1308,"n_targets":22,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1189,"Elo_lo":1091,"Elo_hi":1292,"n_targets":22,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1181,"Elo_lo":1073,"Elo_hi":1301,"n_targets":22,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1155,"Elo_lo":1045,"Elo_hi":1268,"n_targets":22,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1141,"Elo_lo":1032,"Elo_hi":1260,"n_targets":22,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1133,"Elo_lo":1044,"Elo_hi":1224,"n_targets":22,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1111,"Elo_lo":1027,"Elo_hi":1203,"n_targets":22,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1099,"Elo_lo":993,"Elo_hi":1222,"n_targets":14,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1098,"Elo_lo":994,"Elo_hi":1218,"n_targets":14,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1069,"Elo_lo":1001,"Elo_hi":1143,"n_targets":22,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1056,"Elo_lo":955,"Elo_hi":1162,"n_targets":22,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1041,"Elo_lo":965,"Elo_hi":1120,"n_targets":22,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1039,"Elo_lo":949,"Elo_hi":1123,"n_targets":22,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1036,"Elo_lo":902,"Elo_hi":1181,"n_targets":22,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1027,"Elo_lo":943,"Elo_hi":1113,"n_targets":22,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1021,"Elo_lo":941,"Elo_hi":1104,"n_targets":22,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1016,"Elo_lo":919,"Elo_hi":1106,"n_targets":22,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1011,"Elo_lo":881,"Elo_hi":1146,"n_targets":14,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1010,"Elo_lo":960,"Elo_hi":1059,"n_targets":22,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1008,"Elo_lo":867,"Elo_hi":1149,"n_targets":22,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1006,"Elo_lo":977,"Elo_hi":1035,"n_targets":22,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT","Elo":1003,"Elo_lo":969,"Elo_hi":1036,"n_targets":22,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":22,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":986,"Elo_lo":902,"Elo_hi":1071,"n_targets":22,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":978,"Elo_lo":922,"Elo_hi":1033,"n_targets":22,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":977,"Elo_lo":919,"Elo_hi":1032,"n_targets":22,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":954,"Elo_lo":895,"Elo_hi":1021,"n_targets":22,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":916,"Elo_lo":819,"Elo_hi":1013,"n_targets":22,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR","Elo":901,"Elo_lo":765,"Elo_hi":1015,"n_targets":22,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":897,"Elo_lo":845,"Elo_hi":959,"n_targets":22,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":817,"Elo_lo":654,"Elo_hi":950,"n_targets":22,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":762,"Elo_lo":660,"Elo_hi":860,"n_targets":22,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":411,"Elo_lo":310,"Elo_hi":519,"n_targets":22,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1202,"Elo_lo":1066,"Elo_hi":1358,"n_targets":17,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1196,"Elo_lo":1089,"Elo_hi":1310,"n_targets":17,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1183,"Elo_lo":1057,"Elo_hi":1310,"n_targets":17,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1177,"Elo_lo":1064,"Elo_hi":1305,"n_targets":17,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1160,"Elo_lo":1025,"Elo_hi":1296,"n_targets":17,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1099,"Elo_lo":945,"Elo_hi":1265,"n_targets":17,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1092,"Elo_lo":969,"Elo_hi":1211,"n_targets":10,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1092,"Elo_lo":943,"Elo_hi":1244,"n_targets":17,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1085,"Elo_lo":938,"Elo_hi":1233,"n_targets":10,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1077,"Elo_lo":954,"Elo_hi":1222,"n_targets":17,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1038,"Elo_lo":910,"Elo_hi":1165,"n_targets":17,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1026,"Elo_lo":918,"Elo_hi":1142,"n_targets":17,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1011,"Elo_lo":929,"Elo_hi":1091,"n_targets":17,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":17,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":997,"Elo_lo":858,"Elo_hi":1141,"n_targets":17,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":990,"Elo_lo":919,"Elo_hi":1057,"n_targets":17,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT","Elo":990,"Elo_lo":957,"Elo_hi":1021,"n_targets":17,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":988,"Elo_lo":850,"Elo_hi":1111,"n_targets":17,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":976,"Elo_lo":839,"Elo_hi":1099,"n_targets":17,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":967,"Elo_lo":917,"Elo_hi":1018,"n_targets":17,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":966,"Elo_lo":801,"Elo_hi":1151,"n_targets":17,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":956,"Elo_lo":882,"Elo_hi":1020,"n_targets":17,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":946,"Elo_lo":843,"Elo_hi":1039,"n_targets":17,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":930,"Elo_lo":778,"Elo_hi":1067,"n_targets":17,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":930,"Elo_lo":836,"Elo_hi":1013,"n_targets":17,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":925,"Elo_lo":722,"Elo_hi":1123,"n_targets":17,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":923,"Elo_lo":746,"Elo_hi":1102,"n_targets":10,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":854,"Elo_lo":705,"Elo_hi":985,"n_targets":17,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":838,"Elo_lo":710,"Elo_hi":977,"n_targets":17,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR","Elo":777,"Elo_lo":533,"Elo_hi":971,"n_targets":17,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":737,"Elo_lo":602,"Elo_hi":861,"n_targets":17,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":709,"Elo_lo":588,"Elo_hi":811,"n_targets":17,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":525,"Elo_lo":247,"Elo_hi":757,"n_targets":17,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":364,"Elo_lo":198,"Elo_hi":507,"n_targets":17,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1249,"Elo_lo":1189,"Elo_hi":1309,"n_targets":46,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1240,"Elo_lo":1184,"Elo_hi":1297,"n_targets":46,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1235,"Elo_lo":1174,"Elo_hi":1297,"n_targets":46,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1221,"Elo_lo":1173,"Elo_hi":1270,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1215,"Elo_lo":1155,"Elo_hi":1274,"n_targets":46,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1171,"Elo_lo":1110,"Elo_hi":1235,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1158,"Elo_lo":1091,"Elo_hi":1230,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1154,"Elo_lo":1077,"Elo_hi":1231,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1135,"Elo_lo":1079,"Elo_hi":1196,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1114,"Elo_lo":1061,"Elo_hi":1171,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1113,"Elo_lo":1060,"Elo_hi":1171,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1101,"Elo_lo":1034,"Elo_hi":1171,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1093,"Elo_lo":1048,"Elo_hi":1140,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1080,"Elo_lo":1021,"Elo_hi":1137,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1074,"Elo_lo":1018,"Elo_hi":1127,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1073,"Elo_lo":1016,"Elo_hi":1128,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1070,"Elo_lo":1007,"Elo_hi":1133,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1052,"Elo_lo":1006,"Elo_hi":1099,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1045,"Elo_lo":950,"Elo_hi":1137,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1042,"Elo_lo":1006,"Elo_hi":1076,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1040,"Elo_lo":984,"Elo_hi":1096,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1033,"Elo_lo":961,"Elo_hi":1100,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1029,"Elo_lo":1000,"Elo_hi":1055,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1013,"Elo_lo":966,"Elo_hi":1055,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1003,"Elo_lo":981,"Elo_hi":1026,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR","Elo":995,"Elo_lo":919,"Elo_hi":1062,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT","Elo":992,"Elo_lo":970,"Elo_hi":1013,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":987,"Elo_lo":911,"Elo_hi":1057,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":987,"Elo_lo":944,"Elo_hi":1028,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":976,"Elo_lo":906,"Elo_hi":1039,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":881,"Elo_lo":835,"Elo_hi":929,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":648,"Elo_lo":516,"Elo_hi":761,"n_targets":46,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":483,"Elo_lo":415,"Elo_hi":544,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1147,"Elo_lo":1098,"Elo_hi":1200,"n_targets":45,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1142,"Elo_lo":1081,"Elo_hi":1204,"n_targets":45,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1117,"Elo_lo":1053,"Elo_hi":1185,"n_targets":45,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1098,"Elo_lo":1054,"Elo_hi":1143,"n_targets":45,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1084,"Elo_lo":1026,"Elo_hi":1141,"n_targets":45,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1075,"Elo_lo":1034,"Elo_hi":1120,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1069,"Elo_lo":1021,"Elo_hi":1114,"n_targets":45,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1060,"Elo_lo":1016,"Elo_hi":1103,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1053,"Elo_lo":999,"Elo_hi":1105,"n_targets":45,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1048,"Elo_lo":1006,"Elo_hi":1089,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1009,"Elo_lo":945,"Elo_hi":1073,"n_targets":44,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1006,"Elo_lo":948,"Elo_hi":1067,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT","Elo":1005,"Elo_lo":987,"Elo_hi":1025,"n_targets":45,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1003,"Elo_lo":955,"Elo_hi":1048,"n_targets":45,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":45,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":997,"Elo_lo":969,"Elo_hi":1024,"n_targets":45,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":983,"Elo_lo":918,"Elo_hi":1049,"n_targets":45,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":977,"Elo_lo":920,"Elo_hi":1034,"n_targets":45,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":971,"Elo_lo":909,"Elo_hi":1032,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":969,"Elo_lo":932,"Elo_hi":1010,"n_targets":45,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":964,"Elo_lo":904,"Elo_hi":1020,"n_targets":45,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":958,"Elo_lo":918,"Elo_hi":997,"n_targets":45,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":947,"Elo_lo":899,"Elo_hi":994,"n_targets":45,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":940,"Elo_lo":883,"Elo_hi":995,"n_targets":45,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":923,"Elo_lo":878,"Elo_hi":963,"n_targets":45,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":918,"Elo_lo":870,"Elo_hi":964,"n_targets":45,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":911,"Elo_lo":853,"Elo_hi":969,"n_targets":45,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":906,"Elo_lo":862,"Elo_hi":948,"n_targets":45,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":899,"Elo_lo":851,"Elo_hi":950,"n_targets":45,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":846,"Elo_lo":805,"Elo_hi":886,"n_targets":45,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":808,"Elo_lo":724,"Elo_hi":888,"n_targets":45,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":802,"Elo_lo":752,"Elo_hi":847,"n_targets":45,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":738,"Elo_lo":668,"Elo_hi":801,"n_targets":45,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":734,"Elo_lo":665,"Elo_hi":801,"n_targets":45,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1257,"Elo_lo":1201,"Elo_hi":1316,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1222,"Elo_lo":1169,"Elo_hi":1272,"n_targets":46,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1218,"Elo_lo":1163,"Elo_hi":1274,"n_targets":46,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1213,"Elo_lo":1157,"Elo_hi":1271,"n_targets":46,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1175,"Elo_lo":1122,"Elo_hi":1231,"n_targets":46,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1142,"Elo_lo":1095,"Elo_hi":1187,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1128,"Elo_lo":1083,"Elo_hi":1177,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1087,"Elo_lo":1037,"Elo_hi":1139,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1085,"Elo_lo":1036,"Elo_hi":1134,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1078,"Elo_lo":1029,"Elo_hi":1122,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1065,"Elo_lo":1017,"Elo_hi":1114,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1048,"Elo_lo":977,"Elo_hi":1126,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1034,"Elo_lo":958,"Elo_hi":1112,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1027,"Elo_lo":978,"Elo_hi":1079,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1024,"Elo_lo":974,"Elo_hi":1074,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1016,"Elo_lo":964,"Elo_hi":1069,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1014,"Elo_lo":947,"Elo_hi":1081,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1013,"Elo_lo":959,"Elo_hi":1068,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1013,"Elo_lo":954,"Elo_hi":1071,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1000,"Elo_lo":937,"Elo_hi":1062,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT","Elo":994,"Elo_lo":972,"Elo_hi":1016,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":991,"Elo_lo":937,"Elo_hi":1043,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":990,"Elo_lo":949,"Elo_hi":1033,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":973,"Elo_lo":942,"Elo_hi":1007,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":965,"Elo_lo":924,"Elo_hi":1004,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":964,"Elo_lo":929,"Elo_hi":1002,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":962,"Elo_lo":914,"Elo_hi":1011,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":938,"Elo_lo":886,"Elo_hi":993,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":839,"Elo_lo":800,"Elo_hi":877,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":838,"Elo_lo":798,"Elo_hi":876,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":828,"Elo_lo":778,"Elo_hi":874,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":804,"Elo_lo":728,"Elo_hi":882,"n_targets":46,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":646,"Elo_lo":578,"Elo_hi":708,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1239,"Elo_lo":1182,"Elo_hi":1302,"n_targets":44,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1226,"Elo_lo":1159,"Elo_hi":1290,"n_targets":44,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1210,"Elo_lo":1153,"Elo_hi":1268,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1210,"Elo_lo":1145,"Elo_hi":1280,"n_targets":44,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1199,"Elo_lo":1136,"Elo_hi":1268,"n_targets":44,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1193,"Elo_lo":1151,"Elo_hi":1237,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1158,"Elo_lo":1107,"Elo_hi":1206,"n_targets":44,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1110,"Elo_lo":1032,"Elo_hi":1184,"n_targets":44,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1072,"Elo_lo":1024,"Elo_hi":1119,"n_targets":44,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1063,"Elo_lo":998,"Elo_hi":1125,"n_targets":44,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1060,"Elo_lo":1001,"Elo_hi":1120,"n_targets":43,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1055,"Elo_lo":1002,"Elo_hi":1113,"n_targets":44,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1050,"Elo_lo":975,"Elo_hi":1120,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1046,"Elo_lo":993,"Elo_hi":1098,"n_targets":44,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"AUTOGLUON","Elo":1040,"Elo_lo":992,"Elo_hi":1090,"n_targets":44,"id":"AUTOGLUON","display":"AutoGluon","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"AutoGluon automates tabular learning by training multiple models and combining their predictions in ensembles. Here it runs its native extreme preset within a one-hour budget as a separate AutoML reference.","tuning":null,"category":"AutoML","color":"#111827","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1040,"Elo_lo":984,"Elo_hi":1095,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1040,"Elo_lo":990,"Elo_hi":1092,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1039,"Elo_lo":988,"Elo_hi":1094,"n_targets":44,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1035,"Elo_lo":958,"Elo_hi":1103,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1023,"Elo_lo":942,"Elo_hi":1094,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1018,"Elo_lo":954,"Elo_hi":1085,"n_targets":44,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1014,"Elo_lo":972,"Elo_hi":1051,"n_targets":44,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1011,"Elo_lo":969,"Elo_hi":1061,"n_targets":44,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1008,"Elo_lo":940,"Elo_hi":1071,"n_targets":44,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":991,"Elo_lo":962,"Elo_hi":1023,"n_targets":44,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":980,"Elo_lo":946,"Elo_hi":1017,"n_targets":44,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT","Elo":976,"Elo_lo":952,"Elo_hi":998,"n_targets":44,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":967,"Elo_lo":918,"Elo_hi":1016,"n_targets":44,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":915,"Elo_lo":860,"Elo_hi":966,"n_targets":44,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":909,"Elo_lo":863,"Elo_hi":954,"n_targets":44,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":880,"Elo_lo":832,"Elo_hi":927,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":873,"Elo_lo":821,"Elo_hi":924,"n_targets":44,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":788,"Elo_lo":687,"Elo_hi":876,"n_targets":44,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":558,"Elo_lo":487,"Elo_hi":616,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1261,"Elo_lo":1199,"Elo_hi":1329,"n_targets":36,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1221,"Elo_lo":1169,"Elo_hi":1276,"n_targets":36,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1211,"Elo_lo":1137,"Elo_hi":1288,"n_targets":36,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1211,"Elo_lo":1146,"Elo_hi":1278,"n_targets":36,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1204,"Elo_lo":1134,"Elo_hi":1269,"n_targets":36,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1163,"Elo_lo":1093,"Elo_hi":1241,"n_targets":24,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1162,"Elo_lo":1109,"Elo_hi":1225,"n_targets":36,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1158,"Elo_lo":1077,"Elo_hi":1243,"n_targets":36,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1120,"Elo_lo":1062,"Elo_hi":1185,"n_targets":36,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1101,"Elo_lo":1044,"Elo_hi":1157,"n_targets":36,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1100,"Elo_lo":1010,"Elo_hi":1188,"n_targets":24,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1080,"Elo_lo":999,"Elo_hi":1159,"n_targets":36,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1079,"Elo_lo":1011,"Elo_hi":1153,"n_targets":36,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1078,"Elo_lo":1000,"Elo_hi":1160,"n_targets":36,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1066,"Elo_lo":1003,"Elo_hi":1129,"n_targets":36,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1061,"Elo_lo":995,"Elo_hi":1130,"n_targets":36,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1061,"Elo_lo":996,"Elo_hi":1123,"n_targets":36,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1058,"Elo_lo":980,"Elo_hi":1134,"n_targets":36,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1055,"Elo_lo":965,"Elo_hi":1146,"n_targets":24,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1032,"Elo_lo":999,"Elo_hi":1066,"n_targets":36,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1022,"Elo_lo":960,"Elo_hi":1079,"n_targets":36,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1017,"Elo_lo":977,"Elo_hi":1061,"n_targets":36,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1009,"Elo_lo":971,"Elo_hi":1053,"n_targets":36,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1002,"Elo_lo":945,"Elo_hi":1066,"n_targets":36,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":36,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":996,"Elo_lo":926,"Elo_hi":1065,"n_targets":36,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":986,"Elo_lo":936,"Elo_hi":1041,"n_targets":36,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":983,"Elo_lo":931,"Elo_hi":1038,"n_targets":36,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT","Elo":982,"Elo_lo":959,"Elo_hi":1002,"n_targets":36,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":966,"Elo_lo":899,"Elo_hi":1030,"n_targets":36,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":930,"Elo_lo":871,"Elo_hi":996,"n_targets":36,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":870,"Elo_lo":752,"Elo_hi":974,"n_targets":36,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":845,"Elo_lo":785,"Elo_hi":901,"n_targets":36,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":499,"Elo_lo":428,"Elo_hi":569,"n_targets":36,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1213,"Elo_lo":1115,"Elo_hi":1323,"n_targets":22,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1183,"Elo_lo":1112,"Elo_hi":1267,"n_targets":22,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1157,"Elo_lo":1051,"Elo_hi":1266,"n_targets":22,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1157,"Elo_lo":1054,"Elo_hi":1260,"n_targets":22,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1149,"Elo_lo":1052,"Elo_hi":1249,"n_targets":22,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1137,"Elo_lo":1011,"Elo_hi":1267,"n_targets":14,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1114,"Elo_lo":1027,"Elo_hi":1203,"n_targets":22,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1104,"Elo_lo":1001,"Elo_hi":1214,"n_targets":22,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1088,"Elo_lo":1013,"Elo_hi":1173,"n_targets":22,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1087,"Elo_lo":975,"Elo_hi":1205,"n_targets":14,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1060,"Elo_lo":985,"Elo_hi":1138,"n_targets":22,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1048,"Elo_lo":953,"Elo_hi":1134,"n_targets":22,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1044,"Elo_lo":944,"Elo_hi":1148,"n_targets":22,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1044,"Elo_lo":964,"Elo_hi":1124,"n_targets":22,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1031,"Elo_lo":968,"Elo_hi":1086,"n_targets":22,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1029,"Elo_lo":994,"Elo_hi":1066,"n_targets":22,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1024,"Elo_lo":930,"Elo_hi":1110,"n_targets":22,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1014,"Elo_lo":922,"Elo_hi":1102,"n_targets":22,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1013,"Elo_lo":932,"Elo_hi":1089,"n_targets":22,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1004,"Elo_lo":861,"Elo_hi":1138,"n_targets":22,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":22,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT","Elo":999,"Elo_lo":972,"Elo_hi":1023,"n_targets":22,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":994,"Elo_lo":918,"Elo_hi":1064,"n_targets":22,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":987,"Elo_lo":859,"Elo_hi":1114,"n_targets":14,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":977,"Elo_lo":920,"Elo_hi":1037,"n_targets":22,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":976,"Elo_lo":886,"Elo_hi":1060,"n_targets":22,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR","Elo":945,"Elo_lo":798,"Elo_hi":1068,"n_targets":22,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":939,"Elo_lo":840,"Elo_hi":1032,"n_targets":22,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":928,"Elo_lo":857,"Elo_hi":992,"n_targets":22,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":885,"Elo_lo":793,"Elo_hi":972,"n_targets":22,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":827,"Elo_lo":699,"Elo_hi":958,"n_targets":22,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":801,"Elo_lo":638,"Elo_hi":935,"n_targets":22,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":758,"Elo_lo":651,"Elo_hi":848,"n_targets":22,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":433,"Elo_lo":331,"Elo_hi":538,"n_targets":22,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1201,"Elo_lo":1075,"Elo_hi":1343,"n_targets":17,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1193,"Elo_lo":1118,"Elo_hi":1291,"n_targets":17,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1167,"Elo_lo":1053,"Elo_hi":1280,"n_targets":17,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1127,"Elo_lo":992,"Elo_hi":1269,"n_targets":17,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1121,"Elo_lo":1024,"Elo_hi":1234,"n_targets":17,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1113,"Elo_lo":992,"Elo_hi":1234,"n_targets":10,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1108,"Elo_lo":970,"Elo_hi":1259,"n_targets":17,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1079,"Elo_lo":968,"Elo_hi":1212,"n_targets":17,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1076,"Elo_lo":955,"Elo_hi":1214,"n_targets":17,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1063,"Elo_lo":908,"Elo_hi":1226,"n_targets":10,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1041,"Elo_lo":937,"Elo_hi":1148,"n_targets":17,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1023,"Elo_lo":907,"Elo_hi":1137,"n_targets":17,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT","Elo":1003,"Elo_lo":969,"Elo_hi":1040,"n_targets":17,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":17,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":996,"Elo_lo":911,"Elo_hi":1082,"n_targets":17,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":993,"Elo_lo":953,"Elo_hi":1030,"n_targets":17,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":990,"Elo_lo":915,"Elo_hi":1067,"n_targets":17,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":990,"Elo_lo":864,"Elo_hi":1117,"n_targets":17,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":989,"Elo_lo":845,"Elo_hi":1134,"n_targets":17,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":987,"Elo_lo":932,"Elo_hi":1036,"n_targets":17,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":980,"Elo_lo":897,"Elo_hi":1053,"n_targets":17,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":975,"Elo_lo":842,"Elo_hi":1112,"n_targets":17,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":963,"Elo_lo":823,"Elo_hi":1100,"n_targets":17,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":934,"Elo_lo":850,"Elo_hi":1016,"n_targets":17,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":923,"Elo_lo":728,"Elo_hi":1110,"n_targets":17,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":901,"Elo_lo":710,"Elo_hi":1092,"n_targets":10,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":862,"Elo_lo":741,"Elo_hi":990,"n_targets":17,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":857,"Elo_lo":726,"Elo_hi":1014,"n_targets":17,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":844,"Elo_lo":692,"Elo_hi":977,"n_targets":17,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR","Elo":803,"Elo_lo":554,"Elo_hi":999,"n_targets":17,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":707,"Elo_lo":589,"Elo_hi":811,"n_targets":17,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":697,"Elo_lo":570,"Elo_hi":814,"n_targets":17,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":533,"Elo_lo":278,"Elo_hi":758,"n_targets":17,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":386,"Elo_lo":233,"Elo_hi":518,"n_targets":17,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1231,"Elo_lo":1167,"Elo_hi":1294,"n_targets":46,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1224,"Elo_lo":1160,"Elo_hi":1292,"n_targets":46,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1206,"Elo_lo":1161,"Elo_hi":1252,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1198,"Elo_lo":1136,"Elo_hi":1260,"n_targets":46,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1190,"Elo_lo":1131,"Elo_hi":1252,"n_targets":46,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1176,"Elo_lo":1116,"Elo_hi":1239,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1127,"Elo_lo":1041,"Elo_hi":1209,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1122,"Elo_lo":1063,"Elo_hi":1185,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1115,"Elo_lo":1061,"Elo_hi":1168,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1106,"Elo_lo":1051,"Elo_hi":1165,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1090,"Elo_lo":1030,"Elo_hi":1144,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1087,"Elo_lo":1018,"Elo_hi":1158,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1077,"Elo_lo":1027,"Elo_hi":1126,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1076,"Elo_lo":1016,"Elo_hi":1136,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1068,"Elo_lo":991,"Elo_hi":1147,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1066,"Elo_lo":1020,"Elo_hi":1109,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1064,"Elo_lo":1009,"Elo_hi":1116,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1060,"Elo_lo":997,"Elo_hi":1120,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1055,"Elo_lo":1011,"Elo_hi":1097,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1041,"Elo_lo":1016,"Elo_hi":1068,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1033,"Elo_lo":992,"Elo_hi":1072,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1029,"Elo_lo":965,"Elo_hi":1094,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1020,"Elo_lo":945,"Elo_hi":1092,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1018,"Elo_lo":993,"Elo_hi":1046,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1010,"Elo_lo":936,"Elo_hi":1079,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":1001,"Elo_lo":953,"Elo_hi":1047,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":990,"Elo_lo":920,"Elo_hi":1057,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT","Elo":984,"Elo_lo":962,"Elo_hi":1008,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":920,"Elo_lo":826,"Elo_hi":1001,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":890,"Elo_lo":848,"Elo_hi":934,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":799,"Elo_lo":730,"Elo_hi":864,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":619,"Elo_lo":489,"Elo_hi":730,"n_targets":46,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":487,"Elo_lo":421,"Elo_hi":546,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1126,"Elo_lo":1064,"Elo_hi":1189,"n_targets":45,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1110,"Elo_lo":1045,"Elo_hi":1180,"n_targets":45,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1080,"Elo_lo":1034,"Elo_hi":1130,"n_targets":45,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1053,"Elo_lo":1002,"Elo_hi":1104,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1053,"Elo_lo":1008,"Elo_hi":1099,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1040,"Elo_lo":992,"Elo_hi":1088,"n_targets":45,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1040,"Elo_lo":995,"Elo_hi":1085,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1034,"Elo_lo":980,"Elo_hi":1084,"n_targets":45,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1028,"Elo_lo":971,"Elo_hi":1083,"n_targets":45,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":45,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":999,"Elo_lo":934,"Elo_hi":1064,"n_targets":44,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":994,"Elo_lo":922,"Elo_hi":1067,"n_targets":45,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT","Elo":991,"Elo_lo":972,"Elo_hi":1008,"n_targets":45,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":986,"Elo_lo":928,"Elo_hi":1047,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":985,"Elo_lo":933,"Elo_hi":1035,"n_targets":45,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":977,"Elo_lo":917,"Elo_hi":1033,"n_targets":45,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":976,"Elo_lo":913,"Elo_hi":1037,"n_targets":45,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":970,"Elo_lo":935,"Elo_hi":1000,"n_targets":45,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":953,"Elo_lo":890,"Elo_hi":1016,"n_targets":45,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":949,"Elo_lo":908,"Elo_hi":990,"n_targets":45,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":947,"Elo_lo":886,"Elo_hi":1012,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":934,"Elo_lo":883,"Elo_hi":983,"n_targets":45,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":929,"Elo_lo":885,"Elo_hi":973,"n_targets":45,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":921,"Elo_lo":864,"Elo_hi":976,"n_targets":45,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":911,"Elo_lo":863,"Elo_hi":952,"n_targets":45,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":899,"Elo_lo":850,"Elo_hi":945,"n_targets":45,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":889,"Elo_lo":845,"Elo_hi":932,"n_targets":45,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":881,"Elo_lo":819,"Elo_hi":938,"n_targets":45,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":878,"Elo_lo":828,"Elo_hi":931,"n_targets":45,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":818,"Elo_lo":773,"Elo_hi":861,"n_targets":45,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":797,"Elo_lo":710,"Elo_hi":880,"n_targets":45,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":791,"Elo_lo":738,"Elo_hi":840,"n_targets":45,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":729,"Elo_lo":658,"Elo_hi":796,"n_targets":45,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":726,"Elo_lo":655,"Elo_hi":796,"n_targets":45,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1223,"Elo_lo":1159,"Elo_hi":1286,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1208,"Elo_lo":1157,"Elo_hi":1264,"n_targets":46,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1201,"Elo_lo":1143,"Elo_hi":1258,"n_targets":46,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1140,"Elo_lo":1097,"Elo_hi":1183,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1138,"Elo_lo":1091,"Elo_hi":1180,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1135,"Elo_lo":1084,"Elo_hi":1190,"n_targets":46,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1082,"Elo_lo":1034,"Elo_hi":1128,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1072,"Elo_lo":1022,"Elo_hi":1114,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1071,"Elo_lo":1024,"Elo_hi":1119,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1061,"Elo_lo":1009,"Elo_hi":1113,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1041,"Elo_lo":970,"Elo_hi":1113,"n_targets":46,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1026,"Elo_lo":956,"Elo_hi":1100,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1024,"Elo_lo":951,"Elo_hi":1099,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1020,"Elo_lo":968,"Elo_hi":1070,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1018,"Elo_lo":968,"Elo_hi":1068,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":996,"Elo_lo":936,"Elo_hi":1059,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":995,"Elo_lo":941,"Elo_hi":1054,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":994,"Elo_lo":936,"Elo_hi":1048,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT","Elo":994,"Elo_lo":974,"Elo_hi":1013,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":987,"Elo_lo":935,"Elo_hi":1041,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":983,"Elo_lo":944,"Elo_hi":1024,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":971,"Elo_lo":933,"Elo_hi":1008,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":963,"Elo_lo":932,"Elo_hi":997,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":958,"Elo_lo":925,"Elo_hi":996,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":951,"Elo_lo":905,"Elo_hi":998,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":948,"Elo_lo":897,"Elo_hi":1000,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":946,"Elo_lo":894,"Elo_hi":996,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":939,"Elo_lo":892,"Elo_hi":990,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":852,"Elo_lo":807,"Elo_hi":894,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":844,"Elo_lo":801,"Elo_hi":889,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":841,"Elo_lo":793,"Elo_hi":884,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":801,"Elo_lo":723,"Elo_hi":877,"n_targets":46,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":650,"Elo_lo":589,"Elo_hi":711,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1235,"Elo_lo":1171,"Elo_hi":1307,"n_targets":44,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1218,"Elo_lo":1152,"Elo_hi":1290,"n_targets":44,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1206,"Elo_lo":1138,"Elo_hi":1280,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1192,"Elo_lo":1151,"Elo_hi":1234,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1191,"Elo_lo":1132,"Elo_hi":1255,"n_targets":44,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1173,"Elo_lo":1120,"Elo_hi":1229,"n_targets":44,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1099,"Elo_lo":1027,"Elo_hi":1172,"n_targets":44,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1082,"Elo_lo":1036,"Elo_hi":1127,"n_targets":44,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1075,"Elo_lo":1013,"Elo_hi":1135,"n_targets":43,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1059,"Elo_lo":1001,"Elo_hi":1117,"n_targets":44,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1057,"Elo_lo":1002,"Elo_hi":1111,"n_targets":44,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1052,"Elo_lo":982,"Elo_hi":1119,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1049,"Elo_lo":995,"Elo_hi":1103,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1038,"Elo_lo":984,"Elo_hi":1089,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1034,"Elo_lo":962,"Elo_hi":1104,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1033,"Elo_lo":989,"Elo_hi":1075,"n_targets":44,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1031,"Elo_lo":950,"Elo_hi":1107,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1030,"Elo_lo":968,"Elo_hi":1088,"n_targets":44,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1029,"Elo_lo":945,"Elo_hi":1115,"n_targets":44,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1028,"Elo_lo":961,"Elo_hi":1089,"n_targets":44,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1015,"Elo_lo":970,"Elo_hi":1062,"n_targets":44,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1006,"Elo_lo":977,"Elo_hi":1037,"n_targets":44,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":996,"Elo_lo":933,"Elo_hi":1060,"n_targets":44,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":995,"Elo_lo":931,"Elo_hi":1052,"n_targets":44,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT","Elo":988,"Elo_lo":972,"Elo_hi":1003,"n_targets":44,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":987,"Elo_lo":954,"Elo_hi":1020,"n_targets":44,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":930,"Elo_lo":878,"Elo_hi":976,"n_targets":44,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":926,"Elo_lo":871,"Elo_hi":980,"n_targets":44,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":917,"Elo_lo":876,"Elo_hi":957,"n_targets":44,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":889,"Elo_lo":843,"Elo_hi":929,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":888,"Elo_lo":833,"Elo_hi":943,"n_targets":44,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":798,"Elo_lo":704,"Elo_hi":882,"n_targets":44,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":589,"Elo_lo":529,"Elo_hi":642,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1257,"Elo_lo":1198,"Elo_hi":1323,"n_targets":36,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1202,"Elo_lo":1135,"Elo_hi":1265,"n_targets":36,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1199,"Elo_lo":1140,"Elo_hi":1260,"n_targets":36,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1194,"Elo_lo":1144,"Elo_hi":1249,"n_targets":36,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1169,"Elo_lo":1116,"Elo_hi":1224,"n_targets":36,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1157,"Elo_lo":1094,"Elo_hi":1225,"n_targets":24,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1117,"Elo_lo":1042,"Elo_hi":1203,"n_targets":36,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1110,"Elo_lo":1056,"Elo_hi":1169,"n_targets":36,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1098,"Elo_lo":1046,"Elo_hi":1147,"n_targets":36,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1077,"Elo_lo":998,"Elo_hi":1155,"n_targets":36,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1074,"Elo_lo":990,"Elo_hi":1154,"n_targets":24,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1063,"Elo_lo":1005,"Elo_hi":1121,"n_targets":36,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1056,"Elo_lo":999,"Elo_hi":1119,"n_targets":36,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1055,"Elo_lo":982,"Elo_hi":1127,"n_targets":36,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1055,"Elo_lo":1000,"Elo_hi":1109,"n_targets":36,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1043,"Elo_lo":950,"Elo_hi":1130,"n_targets":36,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1043,"Elo_lo":956,"Elo_hi":1126,"n_targets":24,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1037,"Elo_lo":999,"Elo_hi":1079,"n_targets":36,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1035,"Elo_lo":960,"Elo_hi":1107,"n_targets":36,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1024,"Elo_lo":966,"Elo_hi":1077,"n_targets":36,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1023,"Elo_lo":960,"Elo_hi":1081,"n_targets":36,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1019,"Elo_lo":984,"Elo_hi":1054,"n_targets":36,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1008,"Elo_lo":972,"Elo_hi":1045,"n_targets":36,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1001,"Elo_lo":947,"Elo_hi":1062,"n_targets":36,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":36,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":996,"Elo_lo":943,"Elo_hi":1048,"n_targets":36,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT","Elo":991,"Elo_lo":971,"Elo_hi":1008,"n_targets":36,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":984,"Elo_lo":940,"Elo_hi":1031,"n_targets":36,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":970,"Elo_lo":885,"Elo_hi":1053,"n_targets":36,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":934,"Elo_lo":879,"Elo_hi":994,"n_targets":36,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":923,"Elo_lo":851,"Elo_hi":989,"n_targets":36,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":870,"Elo_lo":759,"Elo_hi":970,"n_targets":36,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":857,"Elo_lo":792,"Elo_hi":918,"n_targets":36,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":543,"Elo_lo":484,"Elo_hi":600,"n_targets":36,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1206,"Elo_lo":1110,"Elo_hi":1318,"n_targets":22,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1170,"Elo_lo":1098,"Elo_hi":1259,"n_targets":22,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1136,"Elo_lo":1037,"Elo_hi":1237,"n_targets":22,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1133,"Elo_lo":1011,"Elo_hi":1258,"n_targets":14,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1120,"Elo_lo":1033,"Elo_hi":1206,"n_targets":22,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1103,"Elo_lo":1001,"Elo_hi":1201,"n_targets":22,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1076,"Elo_lo":1000,"Elo_hi":1163,"n_targets":22,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1071,"Elo_lo":965,"Elo_hi":1174,"n_targets":14,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1065,"Elo_lo":990,"Elo_hi":1142,"n_targets":22,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1050,"Elo_lo":957,"Elo_hi":1140,"n_targets":22,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1046,"Elo_lo":969,"Elo_hi":1123,"n_targets":22,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1034,"Elo_lo":933,"Elo_hi":1145,"n_targets":22,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1029,"Elo_lo":929,"Elo_hi":1137,"n_targets":22,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1027,"Elo_lo":996,"Elo_hi":1060,"n_targets":22,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1022,"Elo_lo":922,"Elo_hi":1121,"n_targets":22,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1020,"Elo_lo":928,"Elo_hi":1105,"n_targets":22,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1016,"Elo_lo":958,"Elo_hi":1070,"n_targets":22,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1005,"Elo_lo":866,"Elo_hi":1136,"n_targets":22,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":22,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT","Elo":999,"Elo_lo":974,"Elo_hi":1022,"n_targets":22,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":995,"Elo_lo":918,"Elo_hi":1067,"n_targets":22,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":995,"Elo_lo":904,"Elo_hi":1086,"n_targets":22,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":991,"Elo_lo":901,"Elo_hi":1075,"n_targets":22,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":985,"Elo_lo":915,"Elo_hi":1054,"n_targets":22,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":974,"Elo_lo":840,"Elo_hi":1110,"n_targets":14,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":970,"Elo_lo":917,"Elo_hi":1025,"n_targets":22,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":939,"Elo_lo":862,"Elo_hi":1012,"n_targets":22,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR","Elo":937,"Elo_lo":798,"Elo_hi":1055,"n_targets":22,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":933,"Elo_lo":836,"Elo_hi":1023,"n_targets":22,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":851,"Elo_lo":759,"Elo_hi":938,"n_targets":22,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":814,"Elo_lo":688,"Elo_hi":945,"n_targets":22,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":802,"Elo_lo":650,"Elo_hi":935,"n_targets":22,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":756,"Elo_lo":652,"Elo_hi":847,"n_targets":22,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":472,"Elo_lo":372,"Elo_hi":564,"n_targets":22,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1206,"Elo_lo":1078,"Elo_hi":1352,"n_targets":17,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1205,"Elo_lo":1126,"Elo_hi":1308,"n_targets":17,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1158,"Elo_lo":1049,"Elo_hi":1274,"n_targets":17,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1115,"Elo_lo":992,"Elo_hi":1238,"n_targets":10,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1112,"Elo_lo":972,"Elo_hi":1269,"n_targets":17,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1111,"Elo_lo":986,"Elo_hi":1245,"n_targets":17,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1091,"Elo_lo":989,"Elo_hi":1216,"n_targets":17,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1083,"Elo_lo":970,"Elo_hi":1216,"n_targets":17,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1082,"Elo_lo":957,"Elo_hi":1222,"n_targets":17,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1062,"Elo_lo":903,"Elo_hi":1231,"n_targets":10,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1051,"Elo_lo":943,"Elo_hi":1162,"n_targets":17,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1015,"Elo_lo":901,"Elo_hi":1126,"n_targets":17,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1001,"Elo_lo":914,"Elo_hi":1087,"n_targets":17,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":17,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT","Elo":997,"Elo_lo":963,"Elo_hi":1035,"n_targets":17,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":993,"Elo_lo":952,"Elo_hi":1031,"n_targets":17,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":985,"Elo_lo":844,"Elo_hi":1129,"n_targets":17,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":984,"Elo_lo":898,"Elo_hi":1057,"n_targets":17,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":980,"Elo_lo":855,"Elo_hi":1113,"n_targets":17,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":979,"Elo_lo":909,"Elo_hi":1046,"n_targets":17,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":978,"Elo_lo":929,"Elo_hi":1025,"n_targets":17,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":970,"Elo_lo":826,"Elo_hi":1110,"n_targets":17,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":958,"Elo_lo":834,"Elo_hi":1083,"n_targets":17,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":930,"Elo_lo":732,"Elo_hi":1122,"n_targets":17,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":926,"Elo_lo":849,"Elo_hi":998,"n_targets":17,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":912,"Elo_lo":711,"Elo_hi":1114,"n_targets":10,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":863,"Elo_lo":742,"Elo_hi":993,"n_targets":17,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":856,"Elo_lo":726,"Elo_hi":1014,"n_targets":17,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":838,"Elo_lo":688,"Elo_hi":966,"n_targets":17,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR","Elo":794,"Elo_lo":553,"Elo_hi":986,"n_targets":17,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":713,"Elo_lo":593,"Elo_hi":817,"n_targets":17,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":700,"Elo_lo":580,"Elo_hi":813,"n_targets":17,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":537,"Elo_lo":288,"Elo_hi":758,"n_targets":17,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":391,"Elo_lo":241,"Elo_hi":523,"n_targets":17,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1196,"Elo_lo":1150,"Elo_hi":1245,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1185,"Elo_lo":1117,"Elo_hi":1251,"n_targets":46,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1177,"Elo_lo":1114,"Elo_hi":1242,"n_targets":46,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1170,"Elo_lo":1112,"Elo_hi":1228,"n_targets":46,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1153,"Elo_lo":1094,"Elo_hi":1213,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1122,"Elo_lo":1065,"Elo_hi":1181,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1107,"Elo_lo":1016,"Elo_hi":1197,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1095,"Elo_lo":1044,"Elo_hi":1145,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1089,"Elo_lo":1032,"Elo_hi":1150,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1075,"Elo_lo":1018,"Elo_hi":1132,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1066,"Elo_lo":1022,"Elo_hi":1111,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1065,"Elo_lo":1010,"Elo_hi":1115,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1061,"Elo_lo":994,"Elo_hi":1127,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1060,"Elo_lo":1021,"Elo_hi":1098,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1046,"Elo_lo":995,"Elo_hi":1095,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1045,"Elo_lo":1000,"Elo_hi":1090,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1041,"Elo_lo":982,"Elo_hi":1098,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1035,"Elo_lo":996,"Elo_hi":1076,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1032,"Elo_lo":957,"Elo_hi":1106,"n_targets":46,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1032,"Elo_lo":999,"Elo_hi":1064,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1015,"Elo_lo":992,"Elo_hi":1041,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1008,"Elo_lo":938,"Elo_hi":1077,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1006,"Elo_lo":933,"Elo_hi":1078,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":995,"Elo_lo":931,"Elo_hi":1059,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT","Elo":987,"Elo_lo":966,"Elo_hi":1007,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":983,"Elo_lo":942,"Elo_hi":1024,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":977,"Elo_lo":911,"Elo_hi":1040,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":977,"Elo_lo":899,"Elo_hi":1055,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":880,"Elo_lo":836,"Elo_hi":924,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":811,"Elo_lo":712,"Elo_hi":897,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":738,"Elo_lo":673,"Elo_hi":801,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":631,"Elo_lo":525,"Elo_hi":731,"n_targets":46,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":523,"Elo_lo":469,"Elo_hi":569,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1128,"Elo_lo":1066,"Elo_hi":1188,"n_targets":43,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1107,"Elo_lo":1041,"Elo_hi":1172,"n_targets":43,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1086,"Elo_lo":1041,"Elo_hi":1135,"n_targets":43,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1044,"Elo_lo":996,"Elo_hi":1092,"n_targets":43,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1038,"Elo_lo":985,"Elo_hi":1084,"n_targets":43,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1028,"Elo_lo":968,"Elo_hi":1085,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1014,"Elo_lo":959,"Elo_hi":1065,"n_targets":43,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1012,"Elo_lo":950,"Elo_hi":1076,"n_targets":43,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":43,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":997,"Elo_lo":929,"Elo_hi":1059,"n_targets":42,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"XT","Elo":994,"Elo_lo":973,"Elo_hi":1015,"n_targets":43,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":992,"Elo_lo":927,"Elo_hi":1053,"n_targets":43,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":991,"Elo_lo":916,"Elo_hi":1068,"n_targets":43,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":984,"Elo_lo":927,"Elo_hi":1047,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":983,"Elo_lo":923,"Elo_hi":1045,"n_targets":43,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":977,"Elo_lo":922,"Elo_hi":1036,"n_targets":43,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":965,"Elo_lo":911,"Elo_hi":1019,"n_targets":43,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":962,"Elo_lo":925,"Elo_hi":999,"n_targets":43,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":952,"Elo_lo":890,"Elo_hi":1009,"n_targets":43,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":948,"Elo_lo":899,"Elo_hi":996,"n_targets":43,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":942,"Elo_lo":877,"Elo_hi":1004,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":933,"Elo_lo":885,"Elo_hi":980,"n_targets":43,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":927,"Elo_lo":871,"Elo_hi":975,"n_targets":43,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":923,"Elo_lo":865,"Elo_hi":978,"n_targets":43,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":912,"Elo_lo":862,"Elo_hi":956,"n_targets":43,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":905,"Elo_lo":854,"Elo_hi":951,"n_targets":43,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":899,"Elo_lo":856,"Elo_hi":939,"n_targets":43,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":884,"Elo_lo":832,"Elo_hi":936,"n_targets":43,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":883,"Elo_lo":819,"Elo_hi":945,"n_targets":43,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":836,"Elo_lo":789,"Elo_hi":881,"n_targets":43,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":800,"Elo_lo":711,"Elo_hi":886,"n_targets":43,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":796,"Elo_lo":743,"Elo_hi":848,"n_targets":43,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":734,"Elo_lo":667,"Elo_hi":801,"n_targets":43,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":729,"Elo_lo":661,"Elo_hi":794,"n_targets":43,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1202,"Elo_lo":1145,"Elo_hi":1255,"n_targets":44,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1198,"Elo_lo":1138,"Elo_hi":1255,"n_targets":44,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1183,"Elo_lo":1112,"Elo_hi":1253,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1133,"Elo_lo":1090,"Elo_hi":1176,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1131,"Elo_lo":1077,"Elo_hi":1182,"n_targets":44,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1082,"Elo_lo":1030,"Elo_hi":1136,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1079,"Elo_lo":1020,"Elo_hi":1132,"n_targets":44,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1058,"Elo_lo":1010,"Elo_hi":1102,"n_targets":44,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1056,"Elo_lo":1002,"Elo_hi":1104,"n_targets":43,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1046,"Elo_lo":971,"Elo_hi":1125,"n_targets":44,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1024,"Elo_lo":949,"Elo_hi":1091,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1022,"Elo_lo":950,"Elo_hi":1100,"n_targets":44,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1012,"Elo_lo":955,"Elo_hi":1070,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1004,"Elo_lo":942,"Elo_hi":1064,"n_targets":44,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":998,"Elo_lo":944,"Elo_hi":1050,"n_targets":44,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"XT","Elo":996,"Elo_lo":973,"Elo_hi":1017,"n_targets":44,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":995,"Elo_lo":934,"Elo_hi":1054,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":982,"Elo_lo":922,"Elo_hi":1045,"n_targets":44,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":978,"Elo_lo":939,"Elo_hi":1018,"n_targets":44,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":978,"Elo_lo":940,"Elo_hi":1014,"n_targets":44,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":968,"Elo_lo":937,"Elo_hi":1000,"n_targets":44,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":968,"Elo_lo":907,"Elo_hi":1024,"n_targets":44,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":967,"Elo_lo":909,"Elo_hi":1027,"n_targets":44,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":958,"Elo_lo":910,"Elo_hi":1008,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":952,"Elo_lo":898,"Elo_hi":1004,"n_targets":44,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":952,"Elo_lo":913,"Elo_hi":987,"n_targets":44,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":933,"Elo_lo":883,"Elo_hi":988,"n_targets":44,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":931,"Elo_lo":880,"Elo_hi":980,"n_targets":44,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":862,"Elo_lo":816,"Elo_hi":907,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":855,"Elo_lo":806,"Elo_hi":901,"n_targets":44,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":845,"Elo_lo":808,"Elo_hi":884,"n_targets":44,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":822,"Elo_lo":743,"Elo_hi":900,"n_targets":44,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":675,"Elo_lo":614,"Elo_hi":732,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1196,"Elo_lo":1121,"Elo_hi":1266,"n_targets":42,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1182,"Elo_lo":1107,"Elo_hi":1253,"n_targets":42,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1179,"Elo_lo":1136,"Elo_hi":1221,"n_targets":42,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1153,"Elo_lo":1078,"Elo_hi":1232,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1147,"Elo_lo":1084,"Elo_hi":1214,"n_targets":42,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1067,"Elo_lo":992,"Elo_hi":1137,"n_targets":42,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1058,"Elo_lo":998,"Elo_hi":1118,"n_targets":41,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1044,"Elo_lo":986,"Elo_hi":1097,"n_targets":42,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1041,"Elo_lo":985,"Elo_hi":1095,"n_targets":42,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1037,"Elo_lo":973,"Elo_hi":1098,"n_targets":42,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1035,"Elo_lo":943,"Elo_hi":1122,"n_targets":42,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1020,"Elo_lo":964,"Elo_hi":1078,"n_targets":42,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1020,"Elo_lo":942,"Elo_hi":1095,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1019,"Elo_lo":946,"Elo_hi":1085,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1018,"Elo_lo":965,"Elo_hi":1071,"n_targets":42,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1012,"Elo_lo":965,"Elo_hi":1053,"n_targets":42,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1008,"Elo_lo":927,"Elo_hi":1090,"n_targets":42,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":42,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":998,"Elo_lo":936,"Elo_hi":1059,"n_targets":42,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":995,"Elo_lo":924,"Elo_hi":1061,"n_targets":42,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":987,"Elo_lo":940,"Elo_hi":1034,"n_targets":42,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"XT","Elo":985,"Elo_lo":969,"Elo_hi":1003,"n_targets":42,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":982,"Elo_lo":949,"Elo_hi":1016,"n_targets":42,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":979,"Elo_lo":944,"Elo_hi":1013,"n_targets":42,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":970,"Elo_lo":900,"Elo_hi":1035,"n_targets":42,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":966,"Elo_lo":910,"Elo_hi":1024,"n_targets":42,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":947,"Elo_lo":873,"Elo_hi":1022,"n_targets":42,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":918,"Elo_lo":872,"Elo_hi":964,"n_targets":42,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":900,"Elo_lo":860,"Elo_hi":937,"n_targets":42,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":881,"Elo_lo":820,"Elo_hi":943,"n_targets":42,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":876,"Elo_lo":832,"Elo_hi":919,"n_targets":42,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":872,"Elo_lo":819,"Elo_hi":924,"n_targets":42,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":805,"Elo_lo":718,"Elo_hi":887,"n_targets":42,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":615,"Elo_lo":562,"Elo_hi":661,"n_targets":42,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1246,"Elo_lo":1185,"Elo_hi":1316,"n_targets":34,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1196,"Elo_lo":1144,"Elo_hi":1249,"n_targets":34,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1190,"Elo_lo":1129,"Elo_hi":1248,"n_targets":34,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1182,"Elo_lo":1117,"Elo_hi":1242,"n_targets":34,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1117,"Elo_lo":1035,"Elo_hi":1195,"n_targets":24,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1105,"Elo_lo":1028,"Elo_hi":1178,"n_targets":34,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1077,"Elo_lo":989,"Elo_hi":1167,"n_targets":34,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1072,"Elo_lo":986,"Elo_hi":1147,"n_targets":34,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1068,"Elo_lo":1010,"Elo_hi":1129,"n_targets":34,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1068,"Elo_lo":986,"Elo_hi":1148,"n_targets":24,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1063,"Elo_lo":971,"Elo_hi":1156,"n_targets":34,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1058,"Elo_lo":997,"Elo_hi":1119,"n_targets":34,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1053,"Elo_lo":988,"Elo_hi":1120,"n_targets":34,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1047,"Elo_lo":989,"Elo_hi":1103,"n_targets":34,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1044,"Elo_lo":968,"Elo_hi":1114,"n_targets":34,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1038,"Elo_lo":958,"Elo_hi":1110,"n_targets":34,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1037,"Elo_lo":950,"Elo_hi":1119,"n_targets":24,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1027,"Elo_lo":953,"Elo_hi":1101,"n_targets":34,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1019,"Elo_lo":949,"Elo_hi":1081,"n_targets":34,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1008,"Elo_lo":939,"Elo_hi":1076,"n_targets":34,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1005,"Elo_lo":945,"Elo_hi":1063,"n_targets":34,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":34,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":999,"Elo_lo":959,"Elo_hi":1040,"n_targets":34,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":996,"Elo_lo":948,"Elo_hi":1044,"n_targets":34,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":993,"Elo_lo":939,"Elo_hi":1047,"n_targets":34,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"XT","Elo":992,"Elo_lo":974,"Elo_hi":1009,"n_targets":34,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":990,"Elo_lo":949,"Elo_hi":1030,"n_targets":34,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":988,"Elo_lo":902,"Elo_hi":1077,"n_targets":34,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":962,"Elo_lo":905,"Elo_hi":1024,"n_targets":34,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":955,"Elo_lo":900,"Elo_hi":1010,"n_targets":34,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":940,"Elo_lo":865,"Elo_hi":1013,"n_targets":34,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":887,"Elo_lo":778,"Elo_hi":981,"n_targets":34,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":863,"Elo_lo":799,"Elo_hi":923,"n_targets":34,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":581,"Elo_lo":513,"Elo_hi":636,"n_targets":34,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1201,"Elo_lo":1103,"Elo_hi":1313,"n_targets":22,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1160,"Elo_lo":1091,"Elo_hi":1248,"n_targets":22,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1133,"Elo_lo":1036,"Elo_hi":1235,"n_targets":22,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1096,"Elo_lo":998,"Elo_hi":1193,"n_targets":22,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1096,"Elo_lo":1000,"Elo_hi":1192,"n_targets":22,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1082,"Elo_lo":954,"Elo_hi":1215,"n_targets":14,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1071,"Elo_lo":967,"Elo_hi":1177,"n_targets":14,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1060,"Elo_lo":985,"Elo_hi":1135,"n_targets":22,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1044,"Elo_lo":966,"Elo_hi":1116,"n_targets":22,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1034,"Elo_lo":939,"Elo_hi":1123,"n_targets":22,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1033,"Elo_lo":947,"Elo_hi":1134,"n_targets":22,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1031,"Elo_lo":931,"Elo_hi":1142,"n_targets":22,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1028,"Elo_lo":933,"Elo_hi":1114,"n_targets":22,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1020,"Elo_lo":921,"Elo_hi":1126,"n_targets":22,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1019,"Elo_lo":923,"Elo_hi":1112,"n_targets":22,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1011,"Elo_lo":961,"Elo_hi":1053,"n_targets":22,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1009,"Elo_lo":943,"Elo_hi":1067,"n_targets":22,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1001,"Elo_lo":863,"Elo_hi":1128,"n_targets":22,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":22,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"XT","Elo":996,"Elo_lo":973,"Elo_hi":1019,"n_targets":22,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":983,"Elo_lo":910,"Elo_hi":1052,"n_targets":22,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":982,"Elo_lo":896,"Elo_hi":1065,"n_targets":22,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":977,"Elo_lo":924,"Elo_hi":1033,"n_targets":22,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":973,"Elo_lo":840,"Elo_hi":1106,"n_targets":14,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":970,"Elo_lo":873,"Elo_hi":1063,"n_targets":22,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":952,"Elo_lo":881,"Elo_hi":1022,"n_targets":22,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":937,"Elo_lo":861,"Elo_hi":1010,"n_targets":22,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"LR","Elo":927,"Elo_lo":793,"Elo_hi":1042,"n_targets":22,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":927,"Elo_lo":831,"Elo_hi":1014,"n_targets":22,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":867,"Elo_lo":782,"Elo_hi":953,"n_targets":22,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":814,"Elo_lo":693,"Elo_hi":945,"n_targets":22,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":802,"Elo_lo":656,"Elo_hi":931,"n_targets":22,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":768,"Elo_lo":664,"Elo_hi":856,"n_targets":22,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":490,"Elo_lo":386,"Elo_hi":582,"n_targets":22,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1206,"Elo_lo":1077,"Elo_hi":1351,"n_targets":17,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1205,"Elo_lo":1126,"Elo_hi":1308,"n_targets":17,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1158,"Elo_lo":1048,"Elo_hi":1273,"n_targets":17,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1117,"Elo_lo":994,"Elo_hi":1241,"n_targets":10,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1111,"Elo_lo":972,"Elo_hi":1268,"n_targets":17,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1110,"Elo_lo":986,"Elo_hi":1243,"n_targets":17,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1083,"Elo_lo":970,"Elo_hi":1217,"n_targets":17,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1082,"Elo_lo":957,"Elo_hi":1222,"n_targets":17,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1076,"Elo_lo":971,"Elo_hi":1206,"n_targets":17,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1064,"Elo_lo":905,"Elo_hi":1232,"n_targets":10,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1051,"Elo_lo":943,"Elo_hi":1162,"n_targets":17,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1015,"Elo_lo":901,"Elo_hi":1127,"n_targets":17,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1001,"Elo_lo":914,"Elo_hi":1088,"n_targets":17,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":17,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT","Elo":997,"Elo_lo":963,"Elo_hi":1035,"n_targets":17,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":993,"Elo_lo":952,"Elo_hi":1031,"n_targets":17,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":985,"Elo_lo":845,"Elo_hi":1130,"n_targets":17,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":984,"Elo_lo":898,"Elo_hi":1057,"n_targets":17,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":980,"Elo_lo":855,"Elo_hi":1112,"n_targets":17,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":979,"Elo_lo":909,"Elo_hi":1046,"n_targets":17,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":978,"Elo_lo":929,"Elo_hi":1025,"n_targets":17,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":970,"Elo_lo":826,"Elo_hi":1110,"n_targets":17,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":958,"Elo_lo":834,"Elo_hi":1084,"n_targets":17,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":930,"Elo_lo":732,"Elo_hi":1123,"n_targets":17,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":926,"Elo_lo":849,"Elo_hi":998,"n_targets":17,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":911,"Elo_lo":711,"Elo_hi":1113,"n_targets":10,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":863,"Elo_lo":742,"Elo_hi":992,"n_targets":17,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":857,"Elo_lo":727,"Elo_hi":1014,"n_targets":17,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":839,"Elo_lo":691,"Elo_hi":967,"n_targets":17,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR","Elo":794,"Elo_lo":554,"Elo_hi":986,"n_targets":17,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":714,"Elo_lo":595,"Elo_hi":818,"n_targets":17,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":703,"Elo_lo":585,"Elo_hi":816,"n_targets":17,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":538,"Elo_lo":291,"Elo_hi":759,"n_targets":17,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":393,"Elo_lo":242,"Elo_hi":523,"n_targets":17,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1182,"Elo_lo":1142,"Elo_hi":1228,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1163,"Elo_lo":1094,"Elo_hi":1228,"n_targets":44,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1146,"Elo_lo":1079,"Elo_hi":1207,"n_targets":44,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1145,"Elo_lo":1089,"Elo_hi":1205,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1145,"Elo_lo":1084,"Elo_hi":1210,"n_targets":44,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"AUTOGLUON","Elo":1144,"Elo_lo":1088,"Elo_hi":1200,"n_targets":44,"id":"AUTOGLUON","display":"AutoGluon","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"AutoGluon automates tabular learning by training multiple models and combining their predictions in ensembles. Here it runs its native extreme preset within a one-hour budget as a separate AutoML reference.","tuning":null,"category":"AutoML","color":"#111827","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1068,"Elo_lo":1013,"Elo_hi":1123,"n_targets":44,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1064,"Elo_lo":1019,"Elo_hi":1115,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1047,"Elo_lo":957,"Elo_hi":1131,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1046,"Elo_lo":994,"Elo_hi":1095,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1041,"Elo_lo":999,"Elo_hi":1086,"n_targets":44,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1040,"Elo_lo":995,"Elo_hi":1086,"n_targets":44,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1039,"Elo_lo":966,"Elo_hi":1117,"n_targets":44,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1034,"Elo_lo":961,"Elo_hi":1102,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1032,"Elo_lo":962,"Elo_hi":1100,"n_targets":44,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1031,"Elo_lo":972,"Elo_hi":1087,"n_targets":44,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1018,"Elo_lo":961,"Elo_hi":1072,"n_targets":44,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1005,"Elo_lo":953,"Elo_hi":1056,"n_targets":44,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1005,"Elo_lo":962,"Elo_hi":1046,"n_targets":44,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"LR","Elo":992,"Elo_lo":922,"Elo_hi":1058,"n_targets":44,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":991,"Elo_lo":956,"Elo_hi":1022,"n_targets":44,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":990,"Elo_lo":950,"Elo_hi":1031,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"XT","Elo":987,"Elo_lo":966,"Elo_hi":1006,"n_targets":44,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":986,"Elo_lo":916,"Elo_hi":1052,"n_targets":44,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":985,"Elo_lo":919,"Elo_hi":1052,"n_targets":44,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":984,"Elo_lo":908,"Elo_hi":1057,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":967,"Elo_lo":919,"Elo_hi":1014,"n_targets":44,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":964,"Elo_lo":897,"Elo_hi":1022,"n_targets":43,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":916,"Elo_lo":835,"Elo_hi":997,"n_targets":44,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":876,"Elo_lo":831,"Elo_hi":920,"n_targets":44,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":804,"Elo_lo":719,"Elo_hi":889,"n_targets":44,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":753,"Elo_lo":691,"Elo_hi":810,"n_targets":44,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":654,"Elo_lo":559,"Elo_hi":745,"n_targets":44,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":557,"Elo_lo":503,"Elo_hi":604,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false}]},"adaptive":{"reference":[{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1243,"Elo_lo":1189,"Elo_hi":1306,"n_targets":44,"f1_macro":0.737513,"rmse":1.75962,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1231,"Elo_lo":1165,"Elo_hi":1294,"n_targets":44,"f1_macro":0.740375,"rmse":1.8159,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1215,"Elo_lo":1148,"Elo_hi":1281,"n_targets":44,"f1_macro":0.719262,"rmse":1.7985,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1214,"Elo_lo":1157,"Elo_hi":1272,"n_targets":32,"f1_macro":0.73223,"rmse":null,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1207,"Elo_lo":1145,"Elo_hi":1269,"n_targets":44,"f1_macro":0.718612,"rmse":1.85064,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1195,"Elo_lo":1155,"Elo_hi":1239,"n_targets":46,"f1_macro":0.719085,"rmse":1.70537,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1163,"Elo_lo":1113,"Elo_hi":1211,"n_targets":46,"f1_macro":0.702468,"rmse":1.64802,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1115,"Elo_lo":1044,"Elo_hi":1184,"n_targets":46,"f1_macro":0.672758,"rmse":1.73352,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1076,"Elo_lo":1032,"Elo_hi":1125,"n_targets":46,"f1_macro":0.673752,"rmse":1.7354,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1068,"Elo_lo":1009,"Elo_hi":1127,"n_targets":46,"f1_macro":0.705373,"rmse":1.84672,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1065,"Elo_lo":1004,"Elo_hi":1123,"n_targets":45,"f1_macro":0.684818,"rmse":1.79341,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1061,"Elo_lo":1008,"Elo_hi":1119,"n_targets":46,"f1_macro":0.684705,"rmse":1.85109,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1057,"Elo_lo":989,"Elo_hi":1119,"n_targets":46,"f1_macro":0.691363,"rmse":1.8168,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1052,"Elo_lo":1001,"Elo_hi":1105,"n_targets":46,"f1_macro":0.655681,"rmse":1.7184,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1047,"Elo_lo":993,"Elo_hi":1100,"n_targets":46,"f1_macro":0.68766,"rmse":2.00093,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"AUTOGLUON","Elo":1044,"Elo_lo":997,"Elo_hi":1095,"n_targets":44,"f1_macro":0.684564,"rmse":2.07258,"id":"AUTOGLUON","display":"AutoGluon","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"AutoGluon automates tabular learning by training multiple models and combining their predictions in ensembles. Here it runs its native extreme preset within a one-hour budget as a separate AutoML reference.","tuning":null,"category":"AutoML","color":"#111827","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1044,"Elo_lo":995,"Elo_hi":1096,"n_targets":46,"f1_macro":0.683754,"rmse":1.97959,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1040,"Elo_lo":990,"Elo_hi":1092,"n_targets":46,"f1_macro":0.682015,"rmse":2.0128,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1039,"Elo_lo":966,"Elo_hi":1108,"n_targets":33,"f1_macro":0.662633,"rmse":null,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1025,"Elo_lo":963,"Elo_hi":1084,"n_targets":46,"f1_macro":0.630921,"rmse":1.69689,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1023,"Elo_lo":945,"Elo_hi":1098,"n_targets":33,"f1_macro":0.655548,"rmse":null,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1016,"Elo_lo":977,"Elo_hi":1051,"n_targets":46,"f1_macro":0.662487,"rmse":1.91609,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1014,"Elo_lo":948,"Elo_hi":1076,"n_targets":46,"f1_macro":0.688452,"rmse":2.04661,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1012,"Elo_lo":970,"Elo_hi":1060,"n_targets":46,"f1_macro":0.668179,"rmse":1.90777,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"f1_macro":0.654438,"rmse":1.83402,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":991,"Elo_lo":962,"Elo_hi":1022,"n_targets":46,"f1_macro":0.648116,"rmse":1.89736,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":980,"Elo_lo":946,"Elo_hi":1015,"n_targets":46,"f1_macro":0.634352,"rmse":1.87936,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT","Elo":979,"Elo_lo":955,"Elo_hi":1001,"n_targets":46,"f1_macro":0.634835,"rmse":1.89411,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":975,"Elo_lo":926,"Elo_hi":1023,"n_targets":46,"f1_macro":0.637298,"rmse":1.79505,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":923,"Elo_lo":868,"Elo_hi":973,"n_targets":46,"f1_macro":0.629803,"rmse":1.78556,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":915,"Elo_lo":873,"Elo_hi":961,"n_targets":46,"f1_macro":0.615762,"rmse":1.82869,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":883,"Elo_lo":835,"Elo_hi":927,"n_targets":46,"f1_macro":0.625968,"rmse":2.00707,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":878,"Elo_lo":829,"Elo_hi":930,"n_targets":44,"f1_macro":0.619756,"rmse":2.18334,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":801,"Elo_lo":707,"Elo_hi":883,"n_targets":45,"f1_macro":0.380405,"rmse":2.16615,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":556,"Elo_lo":492,"Elo_hi":614,"n_targets":46,"f1_macro":0.33283,"rmse":2.13599,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false}],"cell_elo":[{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1153,"Elo_lo":1102,"Elo_hi":1208,"n_targets":45,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1144,"Elo_lo":1087,"Elo_hi":1207,"n_targets":45,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1120,"Elo_lo":1062,"Elo_hi":1183,"n_targets":45,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1116,"Elo_lo":1057,"Elo_hi":1174,"n_targets":45,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1108,"Elo_lo":1061,"Elo_hi":1153,"n_targets":45,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1075,"Elo_lo":1032,"Elo_hi":1122,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1068,"Elo_lo":1021,"Elo_hi":1114,"n_targets":45,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1067,"Elo_lo":1022,"Elo_hi":1109,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1041,"Elo_lo":987,"Elo_hi":1091,"n_targets":45,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1039,"Elo_lo":995,"Elo_hi":1082,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1024,"Elo_lo":976,"Elo_hi":1069,"n_targets":45,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1021,"Elo_lo":980,"Elo_hi":1063,"n_targets":45,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1005,"Elo_lo":942,"Elo_hi":1065,"n_targets":44,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1002,"Elo_lo":940,"Elo_hi":1066,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT","Elo":1001,"Elo_lo":975,"Elo_hi":1024,"n_targets":45,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":45,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":995,"Elo_lo":957,"Elo_hi":1025,"n_targets":45,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":991,"Elo_lo":926,"Elo_hi":1060,"n_targets":45,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":968,"Elo_lo":905,"Elo_hi":1030,"n_targets":45,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":963,"Elo_lo":897,"Elo_hi":1030,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":951,"Elo_lo":897,"Elo_hi":1005,"n_targets":45,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":946,"Elo_lo":898,"Elo_hi":994,"n_targets":45,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":936,"Elo_lo":882,"Elo_hi":988,"n_targets":45,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":928,"Elo_lo":881,"Elo_hi":969,"n_targets":45,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":920,"Elo_lo":867,"Elo_hi":963,"n_targets":45,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":910,"Elo_lo":832,"Elo_hi":985,"n_targets":45,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":902,"Elo_lo":842,"Elo_hi":962,"n_targets":45,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":900,"Elo_lo":840,"Elo_hi":958,"n_targets":45,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":891,"Elo_lo":843,"Elo_hi":933,"n_targets":45,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":890,"Elo_lo":838,"Elo_hi":943,"n_targets":45,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":845,"Elo_lo":799,"Elo_hi":887,"n_targets":45,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":805,"Elo_lo":754,"Elo_hi":853,"n_targets":45,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":740,"Elo_lo":667,"Elo_hi":807,"n_targets":45,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":739,"Elo_lo":666,"Elo_hi":808,"n_targets":45,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1239,"Elo_lo":1178,"Elo_hi":1303,"n_targets":46,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1232,"Elo_lo":1179,"Elo_hi":1290,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1228,"Elo_lo":1175,"Elo_hi":1278,"n_targets":46,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1222,"Elo_lo":1162,"Elo_hi":1283,"n_targets":46,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1212,"Elo_lo":1154,"Elo_hi":1270,"n_targets":46,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1143,"Elo_lo":1098,"Elo_hi":1191,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1122,"Elo_lo":1073,"Elo_hi":1170,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1103,"Elo_lo":1060,"Elo_hi":1152,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1090,"Elo_lo":1014,"Elo_hi":1165,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1086,"Elo_lo":1038,"Elo_hi":1133,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1073,"Elo_lo":1024,"Elo_hi":1120,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1069,"Elo_lo":1022,"Elo_hi":1118,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1061,"Elo_lo":1010,"Elo_hi":1108,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1028,"Elo_lo":980,"Elo_hi":1076,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1024,"Elo_lo":973,"Elo_hi":1074,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1021,"Elo_lo":974,"Elo_hi":1066,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1021,"Elo_lo":946,"Elo_hi":1094,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1009,"Elo_lo":943,"Elo_hi":1080,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1002,"Elo_lo":956,"Elo_hi":1044,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":999,"Elo_lo":944,"Elo_hi":1056,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT","Elo":996,"Elo_lo":972,"Elo_hi":1022,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":984,"Elo_lo":938,"Elo_hi":1027,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":983,"Elo_lo":955,"Elo_hi":1010,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":980,"Elo_lo":930,"Elo_hi":1032,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":979,"Elo_lo":916,"Elo_hi":1039,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":974,"Elo_lo":940,"Elo_hi":1007,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":960,"Elo_lo":928,"Elo_hi":992,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":928,"Elo_lo":879,"Elo_hi":977,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":891,"Elo_lo":821,"Elo_hi":961,"n_targets":46,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":844,"Elo_lo":796,"Elo_hi":890,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":837,"Elo_lo":801,"Elo_hi":871,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":789,"Elo_lo":731,"Elo_hi":841,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":649,"Elo_lo":576,"Elo_hi":718,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1279,"Elo_lo":1216,"Elo_hi":1346,"n_targets":44,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1259,"Elo_lo":1188,"Elo_hi":1330,"n_targets":44,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1259,"Elo_lo":1200,"Elo_hi":1321,"n_targets":44,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1248,"Elo_lo":1177,"Elo_hi":1322,"n_targets":44,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1231,"Elo_lo":1161,"Elo_hi":1301,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1217,"Elo_lo":1176,"Elo_hi":1266,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1159,"Elo_lo":1102,"Elo_hi":1214,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1146,"Elo_lo":1076,"Elo_hi":1213,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1110,"Elo_lo":1043,"Elo_hi":1177,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1110,"Elo_lo":1061,"Elo_hi":1165,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1106,"Elo_lo":1047,"Elo_hi":1165,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1098,"Elo_lo":1046,"Elo_hi":1155,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1087,"Elo_lo":1030,"Elo_hi":1144,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1084,"Elo_lo":1012,"Elo_hi":1151,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1083,"Elo_lo":1025,"Elo_hi":1144,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1077,"Elo_lo":1027,"Elo_hi":1132,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1076,"Elo_lo":1020,"Elo_hi":1135,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1059,"Elo_lo":980,"Elo_hi":1134,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1051,"Elo_lo":1000,"Elo_hi":1101,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1044,"Elo_lo":1001,"Elo_hi":1091,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1035,"Elo_lo":981,"Elo_hi":1092,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1033,"Elo_lo":959,"Elo_hi":1107,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1022,"Elo_lo":988,"Elo_hi":1056,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1018,"Elo_lo":952,"Elo_hi":1080,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1002,"Elo_lo":978,"Elo_hi":1025,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT","Elo":990,"Elo_lo":970,"Elo_hi":1009,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":975,"Elo_lo":945,"Elo_hi":1008,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":925,"Elo_lo":871,"Elo_hi":978,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":917,"Elo_lo":846,"Elo_hi":984,"n_targets":45,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":912,"Elo_lo":869,"Elo_hi":956,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":890,"Elo_lo":840,"Elo_hi":939,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":876,"Elo_lo":819,"Elo_hi":939,"n_targets":44,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":574,"Elo_lo":501,"Elo_hi":637,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1295,"Elo_lo":1232,"Elo_hi":1369,"n_targets":36,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1240,"Elo_lo":1174,"Elo_hi":1309,"n_targets":36,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1227,"Elo_lo":1174,"Elo_hi":1285,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1225,"Elo_lo":1156,"Elo_hi":1294,"n_targets":36,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1223,"Elo_lo":1153,"Elo_hi":1297,"n_targets":45,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1207,"Elo_lo":1145,"Elo_hi":1269,"n_targets":36,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1190,"Elo_lo":1119,"Elo_hi":1260,"n_targets":24,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1188,"Elo_lo":1133,"Elo_hi":1249,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1148,"Elo_lo":1072,"Elo_hi":1220,"n_targets":45,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1143,"Elo_lo":1088,"Elo_hi":1206,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1122,"Elo_lo":1045,"Elo_hi":1200,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1121,"Elo_lo":1058,"Elo_hi":1188,"n_targets":45,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1118,"Elo_lo":1056,"Elo_hi":1182,"n_targets":45,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1111,"Elo_lo":1037,"Elo_hi":1182,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1106,"Elo_lo":1045,"Elo_hi":1165,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1094,"Elo_lo":1041,"Elo_hi":1147,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1090,"Elo_lo":1019,"Elo_hi":1161,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1083,"Elo_lo":1025,"Elo_hi":1146,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1083,"Elo_lo":1025,"Elo_hi":1144,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1071,"Elo_lo":1009,"Elo_hi":1133,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1063,"Elo_lo":989,"Elo_hi":1131,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1039,"Elo_lo":977,"Elo_hi":1099,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1037,"Elo_lo":1002,"Elo_hi":1069,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1024,"Elo_lo":971,"Elo_hi":1081,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1002,"Elo_lo":970,"Elo_hi":1031,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":993,"Elo_lo":955,"Elo_hi":1032,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT","Elo":986,"Elo_lo":966,"Elo_hi":1006,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":973,"Elo_lo":925,"Elo_hi":1022,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":970,"Elo_lo":882,"Elo_hi":1058,"n_targets":37,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":951,"Elo_lo":913,"Elo_hi":987,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":915,"Elo_lo":866,"Elo_hi":961,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":849,"Elo_lo":785,"Elo_hi":914,"n_targets":36,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":480,"Elo_lo":404,"Elo_hi":546,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1262,"Elo_lo":1179,"Elo_hi":1360,"n_targets":22,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1236,"Elo_lo":1179,"Elo_hi":1294,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1212,"Elo_lo":1127,"Elo_hi":1305,"n_targets":22,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1199,"Elo_lo":1139,"Elo_hi":1266,"n_targets":45,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1184,"Elo_lo":1104,"Elo_hi":1276,"n_targets":22,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1173,"Elo_lo":1086,"Elo_hi":1273,"n_targets":22,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1160,"Elo_lo":1107,"Elo_hi":1216,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1143,"Elo_lo":1089,"Elo_hi":1202,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1139,"Elo_lo":1076,"Elo_hi":1206,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1127,"Elo_lo":1042,"Elo_hi":1237,"n_targets":14,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1110,"Elo_lo":1021,"Elo_hi":1198,"n_targets":45,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1090,"Elo_lo":1030,"Elo_hi":1152,"n_targets":45,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1085,"Elo_lo":1037,"Elo_hi":1137,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1081,"Elo_lo":992,"Elo_hi":1169,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1070,"Elo_lo":1016,"Elo_hi":1123,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1069,"Elo_lo":1009,"Elo_hi":1124,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1069,"Elo_lo":1017,"Elo_hi":1127,"n_targets":45,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1068,"Elo_lo":1004,"Elo_hi":1129,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1063,"Elo_lo":990,"Elo_hi":1138,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1053,"Elo_lo":988,"Elo_hi":1115,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1028,"Elo_lo":969,"Elo_hi":1083,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1020,"Elo_lo":984,"Elo_hi":1053,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1006,"Elo_lo":965,"Elo_hi":1042,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1000,"Elo_lo":981,"Elo_hi":1020,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":999,"Elo_lo":956,"Elo_hi":1042,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT","Elo":998,"Elo_lo":972,"Elo_hi":1022,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR","Elo":992,"Elo_lo":907,"Elo_hi":1065,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":986,"Elo_lo":921,"Elo_hi":1048,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":959,"Elo_lo":915,"Elo_hi":1004,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":956,"Elo_lo":861,"Elo_hi":1056,"n_targets":25,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":918,"Elo_lo":877,"Elo_hi":956,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":794,"Elo_lo":691,"Elo_hi":891,"n_targets":22,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":388,"Elo_lo":293,"Elo_hi":475,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1268,"Elo_lo":1172,"Elo_hi":1380,"n_targets":17,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1249,"Elo_lo":1156,"Elo_hi":1337,"n_targets":17,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1227,"Elo_lo":1172,"Elo_hi":1285,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1226,"Elo_lo":1136,"Elo_hi":1317,"n_targets":17,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1203,"Elo_lo":1142,"Elo_hi":1270,"n_targets":45,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1168,"Elo_lo":1062,"Elo_hi":1286,"n_targets":17,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1159,"Elo_lo":1053,"Elo_hi":1292,"n_targets":10,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1156,"Elo_lo":1087,"Elo_hi":1228,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1142,"Elo_lo":1071,"Elo_hi":1218,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1127,"Elo_lo":1066,"Elo_hi":1192,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1109,"Elo_lo":1015,"Elo_hi":1204,"n_targets":45,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1084,"Elo_lo":1016,"Elo_hi":1154,"n_targets":45,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1081,"Elo_lo":1021,"Elo_hi":1147,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1077,"Elo_lo":980,"Elo_hi":1177,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1071,"Elo_lo":1013,"Elo_hi":1131,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1057,"Elo_lo":989,"Elo_hi":1121,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1052,"Elo_lo":987,"Elo_hi":1117,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1050,"Elo_lo":969,"Elo_hi":1129,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1047,"Elo_lo":974,"Elo_hi":1116,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1023,"Elo_lo":978,"Elo_hi":1071,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1022,"Elo_lo":983,"Elo_hi":1058,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1000,"Elo_lo":960,"Elo_hi":1038,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":987,"Elo_lo":959,"Elo_hi":1015,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT","Elo":987,"Elo_lo":963,"Elo_hi":1010,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":983,"Elo_lo":912,"Elo_hi":1049,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":982,"Elo_lo":901,"Elo_hi":1060,"n_targets":45,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":980,"Elo_lo":908,"Elo_hi":1051,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":974,"Elo_lo":923,"Elo_hi":1025,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR","Elo":971,"Elo_lo":876,"Elo_hi":1058,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":945,"Elo_lo":899,"Elo_hi":988,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":896,"Elo_lo":798,"Elo_hi":999,"n_targets":20,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":786,"Elo_lo":686,"Elo_hi":883,"n_targets":17,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":390,"Elo_lo":288,"Elo_hi":480,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1252,"Elo_lo":1194,"Elo_hi":1312,"n_targets":46,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1242,"Elo_lo":1185,"Elo_hi":1297,"n_targets":46,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1240,"Elo_lo":1180,"Elo_hi":1302,"n_targets":46,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1222,"Elo_lo":1175,"Elo_hi":1272,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1219,"Elo_lo":1161,"Elo_hi":1277,"n_targets":46,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1172,"Elo_lo":1111,"Elo_hi":1237,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1161,"Elo_lo":1086,"Elo_hi":1235,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1157,"Elo_lo":1088,"Elo_hi":1230,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1135,"Elo_lo":1078,"Elo_hi":1196,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1115,"Elo_lo":1062,"Elo_hi":1172,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1114,"Elo_lo":1061,"Elo_hi":1172,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1105,"Elo_lo":1041,"Elo_hi":1175,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1094,"Elo_lo":1049,"Elo_hi":1141,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1084,"Elo_lo":997,"Elo_hi":1169,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1080,"Elo_lo":1021,"Elo_hi":1138,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1074,"Elo_lo":1017,"Elo_hi":1128,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1074,"Elo_lo":1016,"Elo_hi":1129,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1069,"Elo_lo":1004,"Elo_hi":1134,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1053,"Elo_lo":1006,"Elo_hi":1100,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1041,"Elo_lo":1005,"Elo_hi":1077,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1040,"Elo_lo":982,"Elo_hi":1096,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1036,"Elo_lo":964,"Elo_hi":1102,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1028,"Elo_lo":999,"Elo_hi":1055,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1012,"Elo_lo":966,"Elo_hi":1055,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1002,"Elo_lo":980,"Elo_hi":1025,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR","Elo":993,"Elo_lo":916,"Elo_hi":1061,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT","Elo":991,"Elo_lo":968,"Elo_hi":1013,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":987,"Elo_lo":944,"Elo_hi":1029,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":976,"Elo_lo":894,"Elo_hi":1051,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":972,"Elo_lo":899,"Elo_hi":1037,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":874,"Elo_lo":826,"Elo_hi":924,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":778,"Elo_lo":683,"Elo_hi":865,"n_targets":46,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":427,"Elo_lo":349,"Elo_hi":501,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1147,"Elo_lo":1098,"Elo_hi":1200,"n_targets":45,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1142,"Elo_lo":1081,"Elo_hi":1204,"n_targets":45,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1117,"Elo_lo":1053,"Elo_hi":1185,"n_targets":45,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1098,"Elo_lo":1054,"Elo_hi":1143,"n_targets":45,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1084,"Elo_lo":1026,"Elo_hi":1141,"n_targets":45,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1075,"Elo_lo":1034,"Elo_hi":1120,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1069,"Elo_lo":1021,"Elo_hi":1114,"n_targets":45,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1060,"Elo_lo":1016,"Elo_hi":1103,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1053,"Elo_lo":999,"Elo_hi":1105,"n_targets":45,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1048,"Elo_lo":1006,"Elo_hi":1089,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1009,"Elo_lo":945,"Elo_hi":1073,"n_targets":44,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1006,"Elo_lo":948,"Elo_hi":1067,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT","Elo":1005,"Elo_lo":987,"Elo_hi":1025,"n_targets":45,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1003,"Elo_lo":955,"Elo_hi":1048,"n_targets":45,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":45,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":997,"Elo_lo":969,"Elo_hi":1024,"n_targets":45,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":983,"Elo_lo":918,"Elo_hi":1049,"n_targets":45,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":977,"Elo_lo":920,"Elo_hi":1034,"n_targets":45,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":971,"Elo_lo":909,"Elo_hi":1032,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":969,"Elo_lo":932,"Elo_hi":1010,"n_targets":45,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":964,"Elo_lo":904,"Elo_hi":1020,"n_targets":45,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":958,"Elo_lo":918,"Elo_hi":997,"n_targets":45,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":947,"Elo_lo":899,"Elo_hi":994,"n_targets":45,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":940,"Elo_lo":883,"Elo_hi":995,"n_targets":45,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":923,"Elo_lo":878,"Elo_hi":963,"n_targets":45,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":918,"Elo_lo":870,"Elo_hi":964,"n_targets":45,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":911,"Elo_lo":853,"Elo_hi":969,"n_targets":45,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":906,"Elo_lo":862,"Elo_hi":948,"n_targets":45,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":899,"Elo_lo":851,"Elo_hi":950,"n_targets":45,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":846,"Elo_lo":805,"Elo_hi":886,"n_targets":45,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":808,"Elo_lo":724,"Elo_hi":888,"n_targets":45,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":802,"Elo_lo":752,"Elo_hi":847,"n_targets":45,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":738,"Elo_lo":668,"Elo_hi":801,"n_targets":45,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":734,"Elo_lo":665,"Elo_hi":801,"n_targets":45,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1257,"Elo_lo":1201,"Elo_hi":1316,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1222,"Elo_lo":1169,"Elo_hi":1272,"n_targets":46,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1218,"Elo_lo":1163,"Elo_hi":1274,"n_targets":46,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1213,"Elo_lo":1157,"Elo_hi":1271,"n_targets":46,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1178,"Elo_lo":1126,"Elo_hi":1231,"n_targets":46,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1142,"Elo_lo":1095,"Elo_hi":1187,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1128,"Elo_lo":1083,"Elo_hi":1177,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1087,"Elo_lo":1037,"Elo_hi":1139,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1085,"Elo_lo":1036,"Elo_hi":1134,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1078,"Elo_lo":1029,"Elo_hi":1122,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1066,"Elo_lo":1017,"Elo_hi":1114,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1048,"Elo_lo":977,"Elo_hi":1126,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1034,"Elo_lo":958,"Elo_hi":1112,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1026,"Elo_lo":978,"Elo_hi":1079,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1024,"Elo_lo":974,"Elo_hi":1074,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1016,"Elo_lo":964,"Elo_hi":1069,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1014,"Elo_lo":947,"Elo_hi":1081,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1013,"Elo_lo":959,"Elo_hi":1068,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1013,"Elo_lo":954,"Elo_hi":1071,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1000,"Elo_lo":937,"Elo_hi":1062,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT","Elo":994,"Elo_lo":972,"Elo_hi":1016,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":990,"Elo_lo":937,"Elo_hi":1043,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":990,"Elo_lo":949,"Elo_hi":1033,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":973,"Elo_lo":941,"Elo_hi":1007,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":965,"Elo_lo":924,"Elo_hi":1004,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":964,"Elo_lo":929,"Elo_hi":1002,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":962,"Elo_lo":914,"Elo_hi":1011,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":937,"Elo_lo":886,"Elo_hi":993,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":839,"Elo_lo":800,"Elo_hi":877,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":838,"Elo_lo":797,"Elo_hi":876,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":828,"Elo_lo":778,"Elo_hi":874,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":803,"Elo_lo":728,"Elo_hi":881,"n_targets":46,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":646,"Elo_lo":579,"Elo_hi":708,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1243,"Elo_lo":1189,"Elo_hi":1306,"n_targets":44,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1231,"Elo_lo":1165,"Elo_hi":1294,"n_targets":44,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1215,"Elo_lo":1148,"Elo_hi":1281,"n_targets":44,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1214,"Elo_lo":1157,"Elo_hi":1272,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1207,"Elo_lo":1145,"Elo_hi":1269,"n_targets":44,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1195,"Elo_lo":1155,"Elo_hi":1239,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1163,"Elo_lo":1113,"Elo_hi":1211,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1115,"Elo_lo":1044,"Elo_hi":1184,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1076,"Elo_lo":1032,"Elo_hi":1125,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1068,"Elo_lo":1009,"Elo_hi":1127,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1065,"Elo_lo":1004,"Elo_hi":1123,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1061,"Elo_lo":1008,"Elo_hi":1119,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1057,"Elo_lo":989,"Elo_hi":1119,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1052,"Elo_lo":1001,"Elo_hi":1105,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1047,"Elo_lo":993,"Elo_hi":1100,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"AUTOGLUON","Elo":1044,"Elo_lo":997,"Elo_hi":1095,"n_targets":44,"id":"AUTOGLUON","display":"AutoGluon","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"AutoGluon automates tabular learning by training multiple models and combining their predictions in ensembles. Here it runs its native extreme preset within a one-hour budget as a separate AutoML reference.","tuning":null,"category":"AutoML","color":"#111827","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1044,"Elo_lo":995,"Elo_hi":1096,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1040,"Elo_lo":990,"Elo_hi":1092,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1039,"Elo_lo":966,"Elo_hi":1108,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1025,"Elo_lo":963,"Elo_hi":1084,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1023,"Elo_lo":945,"Elo_hi":1098,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1016,"Elo_lo":977,"Elo_hi":1051,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1014,"Elo_lo":948,"Elo_hi":1076,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1012,"Elo_lo":970,"Elo_hi":1060,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":991,"Elo_lo":962,"Elo_hi":1022,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":980,"Elo_lo":946,"Elo_hi":1015,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT","Elo":979,"Elo_lo":955,"Elo_hi":1001,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":975,"Elo_lo":926,"Elo_hi":1023,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":923,"Elo_lo":868,"Elo_hi":973,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":915,"Elo_lo":873,"Elo_hi":961,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":883,"Elo_lo":835,"Elo_hi":927,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":878,"Elo_lo":829,"Elo_hi":930,"n_targets":44,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":801,"Elo_lo":707,"Elo_hi":883,"n_targets":45,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":556,"Elo_lo":492,"Elo_hi":614,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1275,"Elo_lo":1214,"Elo_hi":1341,"n_targets":36,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1233,"Elo_lo":1186,"Elo_hi":1284,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1225,"Elo_lo":1152,"Elo_hi":1297,"n_targets":36,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1224,"Elo_lo":1161,"Elo_hi":1285,"n_targets":36,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1218,"Elo_lo":1154,"Elo_hi":1284,"n_targets":36,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1182,"Elo_lo":1133,"Elo_hi":1237,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1177,"Elo_lo":1108,"Elo_hi":1245,"n_targets":24,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1171,"Elo_lo":1098,"Elo_hi":1244,"n_targets":45,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1135,"Elo_lo":1082,"Elo_hi":1196,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1124,"Elo_lo":1048,"Elo_hi":1205,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1111,"Elo_lo":1062,"Elo_hi":1165,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1104,"Elo_lo":1032,"Elo_hi":1178,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1096,"Elo_lo":1026,"Elo_hi":1160,"n_targets":45,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1094,"Elo_lo":1032,"Elo_hi":1152,"n_targets":45,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1087,"Elo_lo":1030,"Elo_hi":1146,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1081,"Elo_lo":1012,"Elo_hi":1144,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1081,"Elo_lo":1022,"Elo_hi":1137,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1080,"Elo_lo":1019,"Elo_hi":1146,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1071,"Elo_lo":991,"Elo_hi":1152,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1041,"Elo_lo":989,"Elo_hi":1098,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1038,"Elo_lo":972,"Elo_hi":1102,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1027,"Elo_lo":999,"Elo_hi":1058,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1021,"Elo_lo":989,"Elo_hi":1058,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1011,"Elo_lo":976,"Elo_hi":1048,"n_targets":45,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1007,"Elo_lo":956,"Elo_hi":1062,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":994,"Elo_lo":948,"Elo_hi":1038,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":986,"Elo_lo":944,"Elo_hi":1028,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT","Elo":980,"Elo_lo":957,"Elo_hi":1000,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":969,"Elo_lo":911,"Elo_hi":1025,"n_targets":45,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":940,"Elo_lo":889,"Elo_hi":989,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":889,"Elo_lo":773,"Elo_hi":988,"n_targets":37,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":858,"Elo_lo":799,"Elo_hi":912,"n_targets":36,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":491,"Elo_lo":425,"Elo_hi":552,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1241,"Elo_lo":1162,"Elo_hi":1336,"n_targets":22,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1212,"Elo_lo":1159,"Elo_hi":1264,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1188,"Elo_lo":1108,"Elo_hi":1268,"n_targets":22,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1185,"Elo_lo":1107,"Elo_hi":1276,"n_targets":22,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1164,"Elo_lo":1062,"Elo_hi":1286,"n_targets":14,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1151,"Elo_lo":1091,"Elo_hi":1217,"n_targets":45,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1147,"Elo_lo":1089,"Elo_hi":1204,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1135,"Elo_lo":1052,"Elo_hi":1227,"n_targets":22,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1128,"Elo_lo":1059,"Elo_hi":1196,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1124,"Elo_lo":1072,"Elo_hi":1181,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1083,"Elo_lo":993,"Elo_hi":1167,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1079,"Elo_lo":1011,"Elo_hi":1141,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1070,"Elo_lo":1014,"Elo_hi":1122,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1068,"Elo_lo":1006,"Elo_hi":1126,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1059,"Elo_lo":999,"Elo_hi":1121,"n_targets":45,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1058,"Elo_lo":1007,"Elo_hi":1111,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1044,"Elo_lo":967,"Elo_hi":1116,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1043,"Elo_lo":975,"Elo_hi":1102,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1038,"Elo_lo":974,"Elo_hi":1099,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1032,"Elo_lo":996,"Elo_hi":1068,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1025,"Elo_lo":1000,"Elo_hi":1052,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1022,"Elo_lo":931,"Elo_hi":1098,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1019,"Elo_lo":950,"Elo_hi":1087,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1014,"Elo_lo":974,"Elo_hi":1051,"n_targets":45,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1011,"Elo_lo":964,"Elo_hi":1061,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1007,"Elo_lo":959,"Elo_hi":1051,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT","Elo":986,"Elo_lo":963,"Elo_hi":1010,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":985,"Elo_lo":902,"Elo_hi":1070,"n_targets":45,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":947,"Elo_lo":900,"Elo_hi":992,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":917,"Elo_lo":807,"Elo_hi":1021,"n_targets":24,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":892,"Elo_lo":829,"Elo_hi":952,"n_targets":45,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":791,"Elo_lo":695,"Elo_hi":880,"n_targets":22,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":403,"Elo_lo":317,"Elo_hi":486,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1262,"Elo_lo":1165,"Elo_hi":1372,"n_targets":17,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1229,"Elo_lo":1149,"Elo_hi":1317,"n_targets":17,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1211,"Elo_lo":1165,"Elo_hi":1261,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1192,"Elo_lo":1098,"Elo_hi":1292,"n_targets":17,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1175,"Elo_lo":1059,"Elo_hi":1302,"n_targets":10,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1173,"Elo_lo":1072,"Elo_hi":1285,"n_targets":17,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1153,"Elo_lo":1097,"Elo_hi":1218,"n_targets":45,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1140,"Elo_lo":1076,"Elo_hi":1209,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1126,"Elo_lo":1056,"Elo_hi":1202,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1121,"Elo_lo":1062,"Elo_hi":1184,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1086,"Elo_lo":1034,"Elo_hi":1142,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1083,"Elo_lo":988,"Elo_hi":1177,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1065,"Elo_lo":999,"Elo_hi":1131,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1059,"Elo_lo":990,"Elo_hi":1129,"n_targets":45,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1058,"Elo_lo":986,"Elo_hi":1128,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1055,"Elo_lo":994,"Elo_hi":1113,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1049,"Elo_lo":983,"Elo_hi":1113,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1045,"Elo_lo":1000,"Elo_hi":1089,"n_targets":45,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1032,"Elo_lo":987,"Elo_hi":1077,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1032,"Elo_lo":950,"Elo_hi":1116,"n_targets":40,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1032,"Elo_lo":946,"Elo_hi":1115,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1019,"Elo_lo":990,"Elo_hi":1050,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1013,"Elo_lo":988,"Elo_hi":1040,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1012,"Elo_lo":935,"Elo_hi":1086,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1011,"Elo_lo":968,"Elo_hi":1054,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1009,"Elo_lo":938,"Elo_hi":1079,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1000,"Elo_lo":903,"Elo_hi":1086,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT","Elo":984,"Elo_lo":958,"Elo_hi":1010,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":972,"Elo_lo":924,"Elo_hi":1021,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":866,"Elo_lo":754,"Elo_hi":974,"n_targets":19,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":810,"Elo_lo":740,"Elo_hi":877,"n_targets":45,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":782,"Elo_lo":681,"Elo_hi":879,"n_targets":17,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":392,"Elo_lo":293,"Elo_hi":479,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1239,"Elo_lo":1176,"Elo_hi":1301,"n_targets":46,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1230,"Elo_lo":1164,"Elo_hi":1297,"n_targets":46,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1209,"Elo_lo":1162,"Elo_hi":1256,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1203,"Elo_lo":1142,"Elo_hi":1265,"n_targets":46,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1196,"Elo_lo":1137,"Elo_hi":1255,"n_targets":46,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1179,"Elo_lo":1118,"Elo_hi":1242,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1133,"Elo_lo":1048,"Elo_hi":1212,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1124,"Elo_lo":1064,"Elo_hi":1188,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1117,"Elo_lo":1063,"Elo_hi":1171,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1107,"Elo_lo":1050,"Elo_hi":1167,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1093,"Elo_lo":1024,"Elo_hi":1164,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1091,"Elo_lo":1030,"Elo_hi":1146,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1090,"Elo_lo":1021,"Elo_hi":1164,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1079,"Elo_lo":1027,"Elo_hi":1129,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1077,"Elo_lo":1016,"Elo_hi":1138,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1067,"Elo_lo":1021,"Elo_hi":1112,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1065,"Elo_lo":1010,"Elo_hi":1118,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1060,"Elo_lo":996,"Elo_hi":1122,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1056,"Elo_lo":1011,"Elo_hi":1099,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1042,"Elo_lo":1017,"Elo_hi":1070,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1035,"Elo_lo":992,"Elo_hi":1075,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1028,"Elo_lo":960,"Elo_hi":1095,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1023,"Elo_lo":948,"Elo_hi":1098,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1018,"Elo_lo":992,"Elo_hi":1047,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1009,"Elo_lo":934,"Elo_hi":1080,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":1001,"Elo_lo":953,"Elo_hi":1049,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":986,"Elo_lo":914,"Elo_hi":1056,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT","Elo":983,"Elo_lo":960,"Elo_hi":1007,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":976,"Elo_lo":895,"Elo_hi":1052,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":881,"Elo_lo":837,"Elo_hi":928,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":782,"Elo_lo":705,"Elo_hi":852,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":658,"Elo_lo":541,"Elo_hi":760,"n_targets":46,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":437,"Elo_lo":363,"Elo_hi":504,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1126,"Elo_lo":1064,"Elo_hi":1189,"n_targets":45,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1110,"Elo_lo":1045,"Elo_hi":1180,"n_targets":45,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1080,"Elo_lo":1034,"Elo_hi":1130,"n_targets":45,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1053,"Elo_lo":1002,"Elo_hi":1104,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1053,"Elo_lo":1008,"Elo_hi":1099,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1040,"Elo_lo":992,"Elo_hi":1088,"n_targets":45,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1040,"Elo_lo":995,"Elo_hi":1085,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1034,"Elo_lo":980,"Elo_hi":1084,"n_targets":45,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1028,"Elo_lo":971,"Elo_hi":1083,"n_targets":45,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":45,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":999,"Elo_lo":934,"Elo_hi":1064,"n_targets":44,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":994,"Elo_lo":922,"Elo_hi":1067,"n_targets":45,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT","Elo":991,"Elo_lo":972,"Elo_hi":1008,"n_targets":45,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":986,"Elo_lo":928,"Elo_hi":1047,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":985,"Elo_lo":933,"Elo_hi":1035,"n_targets":45,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":977,"Elo_lo":917,"Elo_hi":1033,"n_targets":45,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":976,"Elo_lo":913,"Elo_hi":1037,"n_targets":45,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":970,"Elo_lo":935,"Elo_hi":1000,"n_targets":45,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":953,"Elo_lo":890,"Elo_hi":1016,"n_targets":45,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":949,"Elo_lo":908,"Elo_hi":990,"n_targets":45,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":947,"Elo_lo":886,"Elo_hi":1012,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":934,"Elo_lo":883,"Elo_hi":983,"n_targets":45,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":929,"Elo_lo":885,"Elo_hi":973,"n_targets":45,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":921,"Elo_lo":864,"Elo_hi":976,"n_targets":45,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":911,"Elo_lo":863,"Elo_hi":952,"n_targets":45,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":899,"Elo_lo":850,"Elo_hi":945,"n_targets":45,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":889,"Elo_lo":845,"Elo_hi":932,"n_targets":45,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":881,"Elo_lo":819,"Elo_hi":938,"n_targets":45,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":878,"Elo_lo":828,"Elo_hi":931,"n_targets":45,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":818,"Elo_lo":773,"Elo_hi":861,"n_targets":45,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":797,"Elo_lo":710,"Elo_hi":880,"n_targets":45,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":791,"Elo_lo":738,"Elo_hi":840,"n_targets":45,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":729,"Elo_lo":658,"Elo_hi":796,"n_targets":45,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":726,"Elo_lo":655,"Elo_hi":796,"n_targets":45,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1223,"Elo_lo":1159,"Elo_hi":1286,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1209,"Elo_lo":1157,"Elo_hi":1264,"n_targets":46,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1201,"Elo_lo":1143,"Elo_hi":1259,"n_targets":46,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1140,"Elo_lo":1097,"Elo_hi":1183,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1138,"Elo_lo":1091,"Elo_hi":1180,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1135,"Elo_lo":1084,"Elo_hi":1190,"n_targets":46,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1082,"Elo_lo":1034,"Elo_hi":1129,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1072,"Elo_lo":1022,"Elo_hi":1114,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1071,"Elo_lo":1024,"Elo_hi":1119,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1061,"Elo_lo":1009,"Elo_hi":1113,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1048,"Elo_lo":980,"Elo_hi":1121,"n_targets":46,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1026,"Elo_lo":956,"Elo_hi":1100,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1024,"Elo_lo":951,"Elo_hi":1099,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1020,"Elo_lo":968,"Elo_hi":1070,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1018,"Elo_lo":968,"Elo_hi":1068,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":996,"Elo_lo":936,"Elo_hi":1059,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":995,"Elo_lo":941,"Elo_hi":1054,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":994,"Elo_lo":937,"Elo_hi":1048,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT","Elo":994,"Elo_lo":974,"Elo_hi":1013,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":987,"Elo_lo":934,"Elo_hi":1041,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":983,"Elo_lo":944,"Elo_hi":1024,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":971,"Elo_lo":933,"Elo_hi":1008,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":963,"Elo_lo":932,"Elo_hi":997,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":958,"Elo_lo":925,"Elo_hi":996,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":951,"Elo_lo":906,"Elo_hi":998,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":948,"Elo_lo":897,"Elo_hi":1000,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":946,"Elo_lo":894,"Elo_hi":996,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":939,"Elo_lo":892,"Elo_hi":990,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":852,"Elo_lo":806,"Elo_hi":894,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":844,"Elo_lo":800,"Elo_hi":889,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":841,"Elo_lo":792,"Elo_hi":883,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":801,"Elo_lo":722,"Elo_hi":876,"n_targets":46,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":650,"Elo_lo":589,"Elo_hi":711,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1238,"Elo_lo":1178,"Elo_hi":1311,"n_targets":44,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1223,"Elo_lo":1156,"Elo_hi":1290,"n_targets":44,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1221,"Elo_lo":1150,"Elo_hi":1288,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1195,"Elo_lo":1135,"Elo_hi":1255,"n_targets":44,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1194,"Elo_lo":1155,"Elo_hi":1237,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1178,"Elo_lo":1126,"Elo_hi":1233,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1105,"Elo_lo":1039,"Elo_hi":1171,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1086,"Elo_lo":1042,"Elo_hi":1135,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1078,"Elo_lo":1016,"Elo_hi":1139,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1065,"Elo_lo":1007,"Elo_hi":1121,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1061,"Elo_lo":1010,"Elo_hi":1117,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1059,"Elo_lo":996,"Elo_hi":1117,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1053,"Elo_lo":979,"Elo_hi":1134,"n_targets":44,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1051,"Elo_lo":999,"Elo_hi":1106,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1044,"Elo_lo":993,"Elo_hi":1095,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1039,"Elo_lo":971,"Elo_hi":1106,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1035,"Elo_lo":991,"Elo_hi":1076,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1033,"Elo_lo":969,"Elo_hi":1092,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1031,"Elo_lo":971,"Elo_hi":1085,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1031,"Elo_lo":952,"Elo_hi":1110,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1016,"Elo_lo":972,"Elo_hi":1065,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1005,"Elo_lo":978,"Elo_hi":1034,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1004,"Elo_lo":943,"Elo_hi":1064,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1002,"Elo_lo":943,"Elo_hi":1060,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT","Elo":990,"Elo_lo":974,"Elo_hi":1007,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":985,"Elo_lo":955,"Elo_hi":1018,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":936,"Elo_lo":885,"Elo_hi":982,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":933,"Elo_lo":877,"Elo_hi":986,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":922,"Elo_lo":882,"Elo_hi":961,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":893,"Elo_lo":839,"Elo_hi":948,"n_targets":44,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":890,"Elo_lo":846,"Elo_hi":931,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":811,"Elo_lo":723,"Elo_hi":889,"n_targets":45,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":585,"Elo_lo":526,"Elo_hi":638,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1273,"Elo_lo":1214,"Elo_hi":1337,"n_targets":36,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1218,"Elo_lo":1157,"Elo_hi":1281,"n_targets":36,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1215,"Elo_lo":1159,"Elo_hi":1272,"n_targets":36,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1214,"Elo_lo":1167,"Elo_hi":1266,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1193,"Elo_lo":1143,"Elo_hi":1247,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1173,"Elo_lo":1111,"Elo_hi":1237,"n_targets":24,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1136,"Elo_lo":1067,"Elo_hi":1208,"n_targets":45,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1133,"Elo_lo":1084,"Elo_hi":1190,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1116,"Elo_lo":1070,"Elo_hi":1164,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1099,"Elo_lo":1029,"Elo_hi":1166,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1098,"Elo_lo":1028,"Elo_hi":1173,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1083,"Elo_lo":1034,"Elo_hi":1133,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1082,"Elo_lo":1016,"Elo_hi":1140,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1079,"Elo_lo":1024,"Elo_hi":1138,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1077,"Elo_lo":1026,"Elo_hi":1128,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1069,"Elo_lo":986,"Elo_hi":1149,"n_targets":36,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1065,"Elo_lo":988,"Elo_hi":1139,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1051,"Elo_lo":993,"Elo_hi":1108,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1050,"Elo_lo":997,"Elo_hi":1103,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1045,"Elo_lo":984,"Elo_hi":1106,"n_targets":45,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1038,"Elo_lo":1005,"Elo_hi":1076,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1022,"Elo_lo":996,"Elo_hi":1054,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1017,"Elo_lo":984,"Elo_hi":1048,"n_targets":45,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1009,"Elo_lo":959,"Elo_hi":1061,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1005,"Elo_lo":960,"Elo_hi":1047,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":994,"Elo_lo":919,"Elo_hi":1069,"n_targets":44,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":991,"Elo_lo":953,"Elo_hi":1029,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT","Elo":990,"Elo_lo":971,"Elo_hi":1009,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":939,"Elo_lo":889,"Elo_hi":986,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":932,"Elo_lo":870,"Elo_hi":993,"n_targets":43,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":892,"Elo_lo":788,"Elo_hi":990,"n_targets":37,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":870,"Elo_lo":811,"Elo_hi":928,"n_targets":36,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":534,"Elo_lo":481,"Elo_hi":585,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1238,"Elo_lo":1159,"Elo_hi":1333,"n_targets":22,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1206,"Elo_lo":1151,"Elo_hi":1263,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1172,"Elo_lo":1091,"Elo_hi":1255,"n_targets":22,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1165,"Elo_lo":1060,"Elo_hi":1284,"n_targets":14,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1163,"Elo_lo":1103,"Elo_hi":1220,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1137,"Elo_lo":1057,"Elo_hi":1220,"n_targets":22,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1117,"Elo_lo":1062,"Elo_hi":1176,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1104,"Elo_lo":1039,"Elo_hi":1167,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1095,"Elo_lo":1032,"Elo_hi":1163,"n_targets":43,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1091,"Elo_lo":1023,"Elo_hi":1154,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1074,"Elo_lo":985,"Elo_hi":1156,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1071,"Elo_lo":979,"Elo_hi":1168,"n_targets":22,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1063,"Elo_lo":1000,"Elo_hi":1121,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1062,"Elo_lo":1009,"Elo_hi":1114,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1059,"Elo_lo":1010,"Elo_hi":1108,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1041,"Elo_lo":978,"Elo_hi":1102,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1039,"Elo_lo":979,"Elo_hi":1098,"n_targets":44,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1038,"Elo_lo":955,"Elo_hi":1115,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1031,"Elo_lo":963,"Elo_hi":1091,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1027,"Elo_lo":1006,"Elo_hi":1051,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1025,"Elo_lo":938,"Elo_hi":1101,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1025,"Elo_lo":985,"Elo_hi":1063,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1013,"Elo_lo":977,"Elo_hi":1049,"n_targets":45,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1011,"Elo_lo":961,"Elo_hi":1062,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1010,"Elo_lo":964,"Elo_hi":1058,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1009,"Elo_lo":943,"Elo_hi":1075,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT","Elo":994,"Elo_lo":972,"Elo_hi":1014,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":963,"Elo_lo":880,"Elo_hi":1053,"n_targets":38,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":942,"Elo_lo":898,"Elo_hi":989,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":917,"Elo_lo":810,"Elo_hi":1023,"n_targets":24,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":884,"Elo_lo":815,"Elo_hi":951,"n_targets":37,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":790,"Elo_lo":699,"Elo_hi":874,"n_targets":22,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":444,"Elo_lo":362,"Elo_hi":517,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1262,"Elo_lo":1163,"Elo_hi":1375,"n_targets":17,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1214,"Elo_lo":1134,"Elo_hi":1307,"n_targets":17,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1209,"Elo_lo":1157,"Elo_hi":1264,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1172,"Elo_lo":1069,"Elo_hi":1288,"n_targets":17,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1170,"Elo_lo":1085,"Elo_hi":1264,"n_targets":17,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1169,"Elo_lo":1054,"Elo_hi":1308,"n_targets":10,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1151,"Elo_lo":1085,"Elo_hi":1220,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1143,"Elo_lo":1085,"Elo_hi":1210,"n_targets":39,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1113,"Elo_lo":1055,"Elo_hi":1180,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1106,"Elo_lo":1033,"Elo_hi":1184,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1077,"Elo_lo":1023,"Elo_hi":1134,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1070,"Elo_lo":977,"Elo_hi":1165,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1064,"Elo_lo":992,"Elo_hi":1136,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1046,"Elo_lo":989,"Elo_hi":1101,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1042,"Elo_lo":978,"Elo_hi":1104,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1042,"Elo_lo":974,"Elo_hi":1113,"n_targets":43,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1038,"Elo_lo":998,"Elo_hi":1077,"n_targets":45,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1032,"Elo_lo":970,"Elo_hi":1096,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1026,"Elo_lo":981,"Elo_hi":1070,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1024,"Elo_lo":977,"Elo_hi":1072,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1021,"Elo_lo":934,"Elo_hi":1109,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1016,"Elo_lo":990,"Elo_hi":1043,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1014,"Elo_lo":979,"Elo_hi":1049,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1003,"Elo_lo":919,"Elo_hi":1096,"n_targets":33,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1000,"Elo_lo":930,"Elo_hi":1071,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR","Elo":998,"Elo_lo":903,"Elo_hi":1085,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":996,"Elo_lo":923,"Elo_hi":1068,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT","Elo":986,"Elo_lo":961,"Elo_hi":1010,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":951,"Elo_lo":910,"Elo_hi":993,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":858,"Elo_lo":747,"Elo_hi":965,"n_targets":19,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":824,"Elo_lo":749,"Elo_hi":896,"n_targets":33,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":779,"Elo_lo":682,"Elo_hi":874,"n_targets":17,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":403,"Elo_lo":311,"Elo_hi":486,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1198,"Elo_lo":1151,"Elo_hi":1246,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1192,"Elo_lo":1125,"Elo_hi":1258,"n_targets":46,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1182,"Elo_lo":1119,"Elo_hi":1246,"n_targets":46,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1172,"Elo_lo":1112,"Elo_hi":1231,"n_targets":46,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1155,"Elo_lo":1096,"Elo_hi":1216,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1124,"Elo_lo":1066,"Elo_hi":1182,"n_targets":46,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1116,"Elo_lo":1028,"Elo_hi":1202,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1096,"Elo_lo":1045,"Elo_hi":1147,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1090,"Elo_lo":1033,"Elo_hi":1152,"n_targets":46,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1076,"Elo_lo":1018,"Elo_hi":1134,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1068,"Elo_lo":1023,"Elo_hi":1113,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1066,"Elo_lo":1010,"Elo_hi":1117,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1066,"Elo_lo":1000,"Elo_hi":1132,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1061,"Elo_lo":1021,"Elo_hi":1099,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1047,"Elo_lo":995,"Elo_hi":1096,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1046,"Elo_lo":999,"Elo_hi":1092,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1044,"Elo_lo":973,"Elo_hi":1117,"n_targets":46,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1041,"Elo_lo":980,"Elo_hi":1099,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1036,"Elo_lo":996,"Elo_hi":1077,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1032,"Elo_lo":999,"Elo_hi":1064,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1026,"Elo_lo":958,"Elo_hi":1097,"n_targets":46,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1014,"Elo_lo":991,"Elo_hi":1040,"n_targets":46,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1009,"Elo_lo":934,"Elo_hi":1081,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1007,"Elo_lo":937,"Elo_hi":1078,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":992,"Elo_lo":926,"Elo_hi":1059,"n_targets":46,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT","Elo":987,"Elo_lo":965,"Elo_hi":1007,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":982,"Elo_lo":940,"Elo_hi":1025,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":973,"Elo_lo":904,"Elo_hi":1037,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":873,"Elo_lo":827,"Elo_hi":918,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":869,"Elo_lo":785,"Elo_hi":950,"n_targets":46,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":719,"Elo_lo":644,"Elo_hi":786,"n_targets":46,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":663,"Elo_lo":562,"Elo_hi":757,"n_targets":46,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":483,"Elo_lo":420,"Elo_hi":537,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1128,"Elo_lo":1066,"Elo_hi":1188,"n_targets":43,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1107,"Elo_lo":1041,"Elo_hi":1172,"n_targets":43,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1086,"Elo_lo":1041,"Elo_hi":1135,"n_targets":43,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1044,"Elo_lo":996,"Elo_hi":1092,"n_targets":43,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1038,"Elo_lo":985,"Elo_hi":1084,"n_targets":43,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1028,"Elo_lo":968,"Elo_hi":1085,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1014,"Elo_lo":959,"Elo_hi":1065,"n_targets":43,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1012,"Elo_lo":950,"Elo_hi":1076,"n_targets":43,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":43,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":997,"Elo_lo":929,"Elo_hi":1059,"n_targets":42,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"XT","Elo":994,"Elo_lo":973,"Elo_hi":1015,"n_targets":43,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":992,"Elo_lo":927,"Elo_hi":1053,"n_targets":43,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":991,"Elo_lo":916,"Elo_hi":1068,"n_targets":43,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":984,"Elo_lo":927,"Elo_hi":1047,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":983,"Elo_lo":923,"Elo_hi":1045,"n_targets":43,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":977,"Elo_lo":922,"Elo_hi":1036,"n_targets":43,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":965,"Elo_lo":911,"Elo_hi":1019,"n_targets":43,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":962,"Elo_lo":925,"Elo_hi":999,"n_targets":43,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":952,"Elo_lo":890,"Elo_hi":1009,"n_targets":43,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":948,"Elo_lo":899,"Elo_hi":996,"n_targets":43,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":942,"Elo_lo":877,"Elo_hi":1004,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":933,"Elo_lo":885,"Elo_hi":980,"n_targets":43,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":927,"Elo_lo":871,"Elo_hi":975,"n_targets":43,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":923,"Elo_lo":865,"Elo_hi":978,"n_targets":43,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":912,"Elo_lo":862,"Elo_hi":956,"n_targets":43,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":905,"Elo_lo":854,"Elo_hi":951,"n_targets":43,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":899,"Elo_lo":856,"Elo_hi":939,"n_targets":43,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":884,"Elo_lo":832,"Elo_hi":936,"n_targets":43,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":883,"Elo_lo":819,"Elo_hi":945,"n_targets":43,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":836,"Elo_lo":789,"Elo_hi":881,"n_targets":43,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":800,"Elo_lo":711,"Elo_hi":886,"n_targets":43,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":796,"Elo_lo":743,"Elo_hi":848,"n_targets":43,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":734,"Elo_lo":667,"Elo_hi":801,"n_targets":43,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":729,"Elo_lo":661,"Elo_hi":794,"n_targets":43,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1202,"Elo_lo":1145,"Elo_hi":1256,"n_targets":44,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1198,"Elo_lo":1138,"Elo_hi":1256,"n_targets":44,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1183,"Elo_lo":1112,"Elo_hi":1253,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1133,"Elo_lo":1091,"Elo_hi":1176,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1131,"Elo_lo":1077,"Elo_hi":1182,"n_targets":44,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1082,"Elo_lo":1029,"Elo_hi":1136,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1079,"Elo_lo":1020,"Elo_hi":1131,"n_targets":44,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1058,"Elo_lo":1011,"Elo_hi":1102,"n_targets":44,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1056,"Elo_lo":1002,"Elo_hi":1105,"n_targets":43,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1054,"Elo_lo":980,"Elo_hi":1128,"n_targets":44,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1024,"Elo_lo":948,"Elo_hi":1091,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1022,"Elo_lo":950,"Elo_hi":1100,"n_targets":44,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1012,"Elo_lo":955,"Elo_hi":1071,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1004,"Elo_lo":942,"Elo_hi":1063,"n_targets":44,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":998,"Elo_lo":945,"Elo_hi":1050,"n_targets":44,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"XT","Elo":996,"Elo_lo":973,"Elo_hi":1017,"n_targets":44,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":995,"Elo_lo":934,"Elo_hi":1054,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":985,"Elo_lo":942,"Elo_hi":1029,"n_targets":44,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":982,"Elo_lo":922,"Elo_hi":1046,"n_targets":44,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":978,"Elo_lo":940,"Elo_hi":1014,"n_targets":44,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":968,"Elo_lo":906,"Elo_hi":1023,"n_targets":44,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":967,"Elo_lo":937,"Elo_hi":999,"n_targets":44,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":967,"Elo_lo":909,"Elo_hi":1027,"n_targets":44,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":958,"Elo_lo":909,"Elo_hi":1008,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":952,"Elo_lo":898,"Elo_hi":1004,"n_targets":44,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":952,"Elo_lo":913,"Elo_hi":987,"n_targets":44,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":933,"Elo_lo":882,"Elo_hi":987,"n_targets":44,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":931,"Elo_lo":880,"Elo_hi":980,"n_targets":44,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":862,"Elo_lo":814,"Elo_hi":907,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":854,"Elo_lo":805,"Elo_hi":901,"n_targets":44,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":845,"Elo_lo":807,"Elo_hi":884,"n_targets":44,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":822,"Elo_lo":742,"Elo_hi":899,"n_targets":44,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":674,"Elo_lo":614,"Elo_hi":731,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1215,"Elo_lo":1146,"Elo_hi":1285,"n_targets":42,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1202,"Elo_lo":1130,"Elo_hi":1267,"n_targets":42,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1183,"Elo_lo":1142,"Elo_hi":1226,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1159,"Elo_lo":1080,"Elo_hi":1236,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1154,"Elo_lo":1090,"Elo_hi":1219,"n_targets":42,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1076,"Elo_lo":1004,"Elo_hi":1146,"n_targets":44,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1061,"Elo_lo":999,"Elo_hi":1121,"n_targets":43,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1050,"Elo_lo":994,"Elo_hi":1102,"n_targets":44,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1046,"Elo_lo":959,"Elo_hi":1133,"n_targets":42,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1045,"Elo_lo":981,"Elo_hi":1102,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1043,"Elo_lo":989,"Elo_hi":1096,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1028,"Elo_lo":971,"Elo_hi":1083,"n_targets":44,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1026,"Elo_lo":972,"Elo_hi":1081,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1023,"Elo_lo":955,"Elo_hi":1088,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1019,"Elo_lo":944,"Elo_hi":1094,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1017,"Elo_lo":940,"Elo_hi":1096,"n_targets":44,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1015,"Elo_lo":967,"Elo_hi":1055,"n_targets":44,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":999,"Elo_lo":939,"Elo_hi":1058,"n_targets":44,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":999,"Elo_lo":930,"Elo_hi":1066,"n_targets":44,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":994,"Elo_lo":953,"Elo_hi":1041,"n_targets":44,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"XT","Elo":988,"Elo_lo":972,"Elo_hi":1005,"n_targets":44,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":982,"Elo_lo":949,"Elo_hi":1015,"n_targets":44,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":980,"Elo_lo":912,"Elo_hi":1043,"n_targets":44,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":977,"Elo_lo":942,"Elo_hi":1010,"n_targets":44,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":977,"Elo_lo":917,"Elo_hi":1034,"n_targets":44,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":956,"Elo_lo":886,"Elo_hi":1030,"n_targets":44,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":924,"Elo_lo":878,"Elo_hi":966,"n_targets":44,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":901,"Elo_lo":861,"Elo_hi":936,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":890,"Elo_lo":831,"Elo_hi":950,"n_targets":44,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":884,"Elo_lo":842,"Elo_hi":929,"n_targets":44,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":875,"Elo_lo":822,"Elo_hi":928,"n_targets":42,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":818,"Elo_lo":732,"Elo_hi":897,"n_targets":43,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":610,"Elo_lo":558,"Elo_hi":659,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1261,"Elo_lo":1203,"Elo_hi":1325,"n_targets":34,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1214,"Elo_lo":1168,"Elo_hi":1263,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1204,"Elo_lo":1148,"Elo_hi":1258,"n_targets":34,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1196,"Elo_lo":1138,"Elo_hi":1254,"n_targets":34,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1132,"Elo_lo":1050,"Elo_hi":1212,"n_targets":24,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1130,"Elo_lo":1063,"Elo_hi":1196,"n_targets":40,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1092,"Elo_lo":1011,"Elo_hi":1176,"n_targets":40,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1087,"Elo_lo":1017,"Elo_hi":1153,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1086,"Elo_lo":1015,"Elo_hi":1157,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1084,"Elo_lo":1032,"Elo_hi":1136,"n_targets":43,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1076,"Elo_lo":984,"Elo_hi":1159,"n_targets":34,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1071,"Elo_lo":1015,"Elo_hi":1133,"n_targets":43,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1071,"Elo_lo":1017,"Elo_hi":1122,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1066,"Elo_lo":998,"Elo_hi":1127,"n_targets":44,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1064,"Elo_lo":1015,"Elo_hi":1118,"n_targets":43,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1056,"Elo_lo":979,"Elo_hi":1130,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1055,"Elo_lo":985,"Elo_hi":1124,"n_targets":40,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1045,"Elo_lo":980,"Elo_hi":1108,"n_targets":43,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1040,"Elo_lo":979,"Elo_hi":1098,"n_targets":43,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1027,"Elo_lo":971,"Elo_hi":1085,"n_targets":43,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1014,"Elo_lo":937,"Elo_hi":1089,"n_targets":40,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1009,"Elo_lo":958,"Elo_hi":1060,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1008,"Elo_lo":971,"Elo_hi":1042,"n_targets":43,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1004,"Elo_lo":959,"Elo_hi":1046,"n_targets":43,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":998,"Elo_lo":955,"Elo_hi":1041,"n_targets":44,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":996,"Elo_lo":962,"Elo_hi":1031,"n_targets":42,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"XT","Elo":990,"Elo_lo":971,"Elo_hi":1009,"n_targets":44,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":971,"Elo_lo":919,"Elo_hi":1022,"n_targets":42,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":956,"Elo_lo":911,"Elo_hi":1002,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":949,"Elo_lo":881,"Elo_hi":1014,"n_targets":40,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":908,"Elo_lo":805,"Elo_hi":1001,"n_targets":35,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":873,"Elo_lo":811,"Elo_hi":932,"n_targets":34,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":570,"Elo_lo":511,"Elo_hi":621,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1229,"Elo_lo":1142,"Elo_hi":1319,"n_targets":22,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1192,"Elo_lo":1141,"Elo_hi":1250,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1163,"Elo_lo":1081,"Elo_hi":1246,"n_targets":22,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1137,"Elo_lo":1068,"Elo_hi":1201,"n_targets":37,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1124,"Elo_lo":1042,"Elo_hi":1204,"n_targets":22,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1109,"Elo_lo":996,"Elo_hi":1235,"n_targets":14,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1091,"Elo_lo":1025,"Elo_hi":1155,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1078,"Elo_lo":1012,"Elo_hi":1148,"n_targets":36,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1075,"Elo_lo":1007,"Elo_hi":1145,"n_targets":36,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1067,"Elo_lo":1002,"Elo_hi":1130,"n_targets":42,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1064,"Elo_lo":970,"Elo_hi":1157,"n_targets":22,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1060,"Elo_lo":1009,"Elo_hi":1115,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1058,"Elo_lo":970,"Elo_hi":1142,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1055,"Elo_lo":993,"Elo_hi":1114,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1051,"Elo_lo":1002,"Elo_hi":1102,"n_targets":43,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1038,"Elo_lo":976,"Elo_hi":1098,"n_targets":42,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1026,"Elo_lo":946,"Elo_hi":1107,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1025,"Elo_lo":964,"Elo_hi":1085,"n_targets":43,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1016,"Elo_lo":970,"Elo_hi":1057,"n_targets":42,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1014,"Elo_lo":982,"Elo_hi":1042,"n_targets":43,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1006,"Elo_lo":944,"Elo_hi":1067,"n_targets":41,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1004,"Elo_lo":969,"Elo_hi":1040,"n_targets":41,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1003,"Elo_lo":920,"Elo_hi":1082,"n_targets":44,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":993,"Elo_lo":948,"Elo_hi":1040,"n_targets":44,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":988,"Elo_lo":923,"Elo_hi":1049,"n_targets":43,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"XT","Elo":985,"Elo_lo":963,"Elo_hi":1007,"n_targets":44,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":972,"Elo_lo":924,"Elo_hi":1016,"n_targets":39,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":960,"Elo_lo":870,"Elo_hi":1047,"n_targets":36,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":944,"Elo_lo":897,"Elo_hi":994,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":913,"Elo_lo":798,"Elo_hi":1016,"n_targets":24,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":895,"Elo_lo":832,"Elo_hi":956,"n_targets":36,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":795,"Elo_lo":707,"Elo_hi":873,"n_targets":22,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":459,"Elo_lo":373,"Elo_hi":535,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1252,"Elo_lo":1153,"Elo_hi":1367,"n_targets":17,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1207,"Elo_lo":1120,"Elo_hi":1301,"n_targets":17,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1201,"Elo_lo":1153,"Elo_hi":1255,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1164,"Elo_lo":1040,"Elo_hi":1303,"n_targets":10,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1164,"Elo_lo":1076,"Elo_hi":1260,"n_targets":17,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1161,"Elo_lo":1059,"Elo_hi":1276,"n_targets":17,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1138,"Elo_lo":1071,"Elo_hi":1214,"n_targets":36,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1124,"Elo_lo":1062,"Elo_hi":1197,"n_targets":32,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1113,"Elo_lo":1047,"Elo_hi":1186,"n_targets":33,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1090,"Elo_lo":1013,"Elo_hi":1168,"n_targets":31,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1069,"Elo_lo":1017,"Elo_hi":1128,"n_targets":42,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1058,"Elo_lo":959,"Elo_hi":1157,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1049,"Elo_lo":990,"Elo_hi":1113,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1047,"Elo_lo":974,"Elo_hi":1124,"n_targets":42,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1040,"Elo_lo":966,"Elo_hi":1116,"n_targets":41,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1029,"Elo_lo":990,"Elo_hi":1070,"n_targets":41,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1022,"Elo_lo":958,"Elo_hi":1088,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1021,"Elo_lo":975,"Elo_hi":1068,"n_targets":44,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1021,"Elo_lo":964,"Elo_hi":1088,"n_targets":40,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1015,"Elo_lo":979,"Elo_hi":1053,"n_targets":41,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1013,"Elo_lo":988,"Elo_hi":1039,"n_targets":43,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1007,"Elo_lo":913,"Elo_hi":1102,"n_targets":31,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1007,"Elo_lo":965,"Elo_hi":1050,"n_targets":38,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1000,"Elo_lo":912,"Elo_hi":1094,"n_targets":31,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":986,"Elo_lo":911,"Elo_hi":1062,"n_targets":43,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT","Elo":983,"Elo_lo":958,"Elo_hi":1009,"n_targets":44,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR","Elo":979,"Elo_lo":880,"Elo_hi":1073,"n_targets":44,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":978,"Elo_lo":900,"Elo_hi":1052,"n_targets":43,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":955,"Elo_lo":915,"Elo_hi":998,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":851,"Elo_lo":739,"Elo_hi":959,"n_targets":19,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":826,"Elo_lo":750,"Elo_hi":897,"n_targets":32,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":772,"Elo_lo":675,"Elo_hi":863,"n_targets":17,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":408,"Elo_lo":312,"Elo_hi":491,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1184,"Elo_lo":1143,"Elo_hi":1231,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1182,"Elo_lo":1122,"Elo_hi":1247,"n_targets":44,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1164,"Elo_lo":1104,"Elo_hi":1222,"n_targets":44,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1149,"Elo_lo":1092,"Elo_hi":1209,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1149,"Elo_lo":1086,"Elo_hi":1214,"n_targets":44,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"AUTOGLUON","Elo":1146,"Elo_lo":1091,"Elo_hi":1204,"n_targets":44,"id":"AUTOGLUON","display":"AutoGluon","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"AutoGluon automates tabular learning by training multiple models and combining their predictions in ensembles. Here it runs its native extreme preset within a one-hour budget as a separate AutoML reference.","tuning":null,"category":"AutoML","color":"#111827","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1084,"Elo_lo":1033,"Elo_hi":1135,"n_targets":44,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1066,"Elo_lo":1020,"Elo_hi":1118,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1062,"Elo_lo":971,"Elo_hi":1144,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1052,"Elo_lo":983,"Elo_hi":1116,"n_targets":33,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1048,"Elo_lo":977,"Elo_hi":1127,"n_targets":44,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1047,"Elo_lo":994,"Elo_hi":1098,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1043,"Elo_lo":1000,"Elo_hi":1089,"n_targets":44,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1042,"Elo_lo":995,"Elo_hi":1089,"n_targets":44,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1034,"Elo_lo":963,"Elo_hi":1103,"n_targets":44,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1033,"Elo_lo":974,"Elo_hi":1089,"n_targets":44,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1018,"Elo_lo":959,"Elo_hi":1074,"n_targets":44,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1007,"Elo_lo":955,"Elo_hi":1058,"n_targets":44,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1006,"Elo_lo":963,"Elo_hi":1047,"n_targets":44,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":999,"Elo_lo":923,"Elo_hi":1070,"n_targets":33,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"LR","Elo":992,"Elo_lo":922,"Elo_hi":1060,"n_targets":44,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":991,"Elo_lo":955,"Elo_hi":1023,"n_targets":44,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":990,"Elo_lo":950,"Elo_hi":1032,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":988,"Elo_lo":917,"Elo_hi":1055,"n_targets":44,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"XT","Elo":986,"Elo_lo":964,"Elo_hi":1005,"n_targets":44,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":984,"Elo_lo":916,"Elo_hi":1054,"n_targets":44,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":968,"Elo_lo":920,"Elo_hi":1015,"n_targets":44,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":959,"Elo_lo":890,"Elo_hi":1019,"n_targets":43,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":956,"Elo_lo":886,"Elo_hi":1029,"n_targets":44,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":870,"Elo_lo":790,"Elo_hi":946,"n_targets":44,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":867,"Elo_lo":820,"Elo_hi":912,"n_targets":44,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":735,"Elo_lo":665,"Elo_hi":798,"n_targets":44,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":685,"Elo_lo":594,"Elo_hi":772,"n_targets":44,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":524,"Elo_lo":463,"Elo_hi":578,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false}]},"conditional":{"reference":[{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1243,"Elo_lo":1184,"Elo_hi":1308,"n_targets":44,"f1_macro":0.737513,"rmse":1.75962,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1229,"Elo_lo":1160,"Elo_hi":1296,"n_targets":44,"f1_macro":0.740375,"rmse":1.8159,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1226,"Elo_lo":1162,"Elo_hi":1293,"n_targets":43,"f1_macro":0.741836,"rmse":1.7985,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1215,"Elo_lo":1154,"Elo_hi":1280,"n_targets":43,"f1_macro":0.735623,"rmse":1.85064,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1210,"Elo_lo":1151,"Elo_hi":1270,"n_targets":31,"f1_macro":0.735229,"rmse":null,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1197,"Elo_lo":1154,"Elo_hi":1242,"n_targets":44,"f1_macro":0.719811,"rmse":1.80462,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1172,"Elo_lo":1127,"Elo_hi":1217,"n_targets":43,"f1_macro":0.724803,"rmse":1.74289,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1135,"Elo_lo":1059,"Elo_hi":1207,"n_targets":42,"f1_macro":0.705583,"rmse":1.83574,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1126,"Elo_lo":1059,"Elo_hi":1183,"n_targets":13,"f1_macro":0.629461,"rmse":0.964619,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1082,"Elo_lo":1037,"Elo_hi":1125,"n_targets":43,"f1_macro":0.694887,"rmse":1.83726,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1069,"Elo_lo":1013,"Elo_hi":1131,"n_targets":40,"f1_macro":0.696297,"rmse":1.96288,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1067,"Elo_lo":1005,"Elo_hi":1126,"n_targets":43,"f1_macro":0.685468,"rmse":1.89663,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1064,"Elo_lo":998,"Elo_hi":1128,"n_targets":44,"f1_macro":0.707116,"rmse":1.95582,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1056,"Elo_lo":1005,"Elo_hi":1106,"n_targets":43,"f1_macro":0.675328,"rmse":1.8178,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1052,"Elo_lo":974,"Elo_hi":1125,"n_targets":44,"f1_macro":0.692084,"rmse":1.92319,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1048,"Elo_lo":994,"Elo_hi":1106,"n_targets":40,"f1_macro":0.691585,"rmse":2.12524,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1044,"Elo_lo":971,"Elo_hi":1113,"n_targets":31,"f1_macro":0.683262,"rmse":null,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"AUTOGLUON","Elo":1041,"Elo_lo":990,"Elo_hi":1091,"n_targets":44,"f1_macro":0.684564,"rmse":2.07258,"id":"AUTOGLUON","display":"AutoGluon","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"AutoGluon automates tabular learning by training multiple models and combining their predictions in ensembles. Here it runs its native extreme preset within a one-hour budget as a separate AutoML reference.","tuning":null,"category":"AutoML","color":"#111827","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1041,"Elo_lo":984,"Elo_hi":1098,"n_targets":44,"f1_macro":0.686295,"rmse":2.11554,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1041,"Elo_lo":988,"Elo_hi":1094,"n_targets":44,"f1_macro":0.684965,"rmse":2.09472,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1033,"Elo_lo":991,"Elo_hi":1072,"n_targets":40,"f1_macro":0.679125,"rmse":2.02183,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1029,"Elo_lo":965,"Elo_hi":1095,"n_targets":43,"f1_macro":0.649173,"rmse":1.79114,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1029,"Elo_lo":946,"Elo_hi":1099,"n_targets":31,"f1_macro":0.67627,"rmse":null,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1011,"Elo_lo":969,"Elo_hi":1064,"n_targets":44,"f1_macro":0.668661,"rmse":2.01502,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1007,"Elo_lo":937,"Elo_hi":1073,"n_targets":44,"f1_macro":0.689285,"rmse":2.17055,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1005,"Elo_lo":977,"Elo_hi":1037,"n_targets":40,"f1_macro":0.660046,"rmse":1.99831,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"f1_macro":0.65413,"rmse":1.93045,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":994,"Elo_lo":957,"Elo_hi":1035,"n_targets":42,"f1_macro":0.662478,"rmse":1.97764,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT","Elo":977,"Elo_lo":951,"Elo_hi":999,"n_targets":44,"f1_macro":0.633911,"rmse":1.99815,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":973,"Elo_lo":924,"Elo_hi":1023,"n_targets":43,"f1_macro":0.656607,"rmse":1.90252,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":918,"Elo_lo":869,"Elo_hi":967,"n_targets":40,"f1_macro":0.629057,"rmse":1.92813,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":910,"Elo_lo":853,"Elo_hi":964,"n_targets":44,"f1_macro":0.628504,"rmse":1.88066,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":875,"Elo_lo":823,"Elo_hi":925,"n_targets":44,"f1_macro":0.625905,"rmse":2.10961,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":868,"Elo_lo":813,"Elo_hi":923,"n_targets":44,"f1_macro":0.619756,"rmse":2.18334,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":526,"Elo_lo":434,"Elo_hi":601,"n_targets":44,"f1_macro":0.33199,"rmse":2.23666,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false}],"cell_elo":[{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1155,"Elo_lo":1103,"Elo_hi":1214,"n_targets":45,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1146,"Elo_lo":1086,"Elo_hi":1211,"n_targets":45,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1122,"Elo_lo":1061,"Elo_hi":1187,"n_targets":45,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1117,"Elo_lo":1056,"Elo_hi":1177,"n_targets":45,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1111,"Elo_lo":1063,"Elo_hi":1156,"n_targets":45,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1078,"Elo_lo":1036,"Elo_hi":1119,"n_targets":44,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1074,"Elo_lo":1030,"Elo_hi":1123,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1069,"Elo_lo":1020,"Elo_hi":1116,"n_targets":45,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1050,"Elo_lo":1010,"Elo_hi":1091,"n_targets":44,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1043,"Elo_lo":987,"Elo_hi":1095,"n_targets":45,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1033,"Elo_lo":989,"Elo_hi":1077,"n_targets":44,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1030,"Elo_lo":993,"Elo_hi":1068,"n_targets":44,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1022,"Elo_lo":936,"Elo_hi":1109,"n_targets":29,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1017,"Elo_lo":987,"Elo_hi":1047,"n_targets":30,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1013,"Elo_lo":950,"Elo_hi":1074,"n_targets":44,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1012,"Elo_lo":950,"Elo_hi":1073,"n_targets":31,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1010,"Elo_lo":945,"Elo_hi":1077,"n_targets":43,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1010,"Elo_lo":962,"Elo_hi":1059,"n_targets":30,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT","Elo":1003,"Elo_lo":977,"Elo_hi":1027,"n_targets":45,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":45,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":995,"Elo_lo":956,"Elo_hi":1027,"n_targets":45,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":969,"Elo_lo":902,"Elo_hi":1034,"n_targets":31,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":966,"Elo_lo":901,"Elo_hi":1025,"n_targets":30,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":965,"Elo_lo":898,"Elo_hi":1029,"n_targets":45,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":957,"Elo_lo":912,"Elo_hi":993,"n_targets":30,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":955,"Elo_lo":906,"Elo_hi":1002,"n_targets":44,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":950,"Elo_lo":894,"Elo_hi":1007,"n_targets":45,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":920,"Elo_lo":866,"Elo_hi":966,"n_targets":45,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":899,"Elo_lo":834,"Elo_hi":960,"n_targets":45,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":890,"Elo_lo":836,"Elo_hi":946,"n_targets":45,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":848,"Elo_lo":790,"Elo_hi":902,"n_targets":30,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":838,"Elo_lo":789,"Elo_hi":883,"n_targets":45,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":740,"Elo_lo":656,"Elo_hi":812,"n_targets":45,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n20","cell_label":"p=2,000, n=20","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":739,"Elo_lo":654,"Elo_hi":813,"n_targets":45,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1242,"Elo_lo":1180,"Elo_hi":1308,"n_targets":46,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1232,"Elo_lo":1177,"Elo_hi":1292,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1231,"Elo_lo":1176,"Elo_hi":1284,"n_targets":46,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1225,"Elo_lo":1163,"Elo_hi":1287,"n_targets":46,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1215,"Elo_lo":1156,"Elo_hi":1276,"n_targets":46,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1146,"Elo_lo":1099,"Elo_hi":1194,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1136,"Elo_lo":1089,"Elo_hi":1182,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1112,"Elo_lo":1038,"Elo_hi":1186,"n_targets":44,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1105,"Elo_lo":1061,"Elo_hi":1156,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1096,"Elo_lo":1050,"Elo_hi":1142,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1083,"Elo_lo":1037,"Elo_hi":1131,"n_targets":45,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1076,"Elo_lo":1028,"Elo_hi":1126,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1062,"Elo_lo":1009,"Elo_hi":1111,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1061,"Elo_lo":1013,"Elo_hi":1108,"n_targets":39,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1045,"Elo_lo":976,"Elo_hi":1117,"n_targets":24,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1037,"Elo_lo":989,"Elo_hi":1085,"n_targets":45,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1033,"Elo_lo":981,"Elo_hi":1084,"n_targets":45,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1032,"Elo_lo":974,"Elo_hi":1097,"n_targets":39,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1028,"Elo_lo":948,"Elo_hi":1103,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1022,"Elo_lo":999,"Elo_hi":1047,"n_targets":39,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1018,"Elo_lo":951,"Elo_hi":1087,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1012,"Elo_lo":981,"Elo_hi":1045,"n_targets":39,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1003,"Elo_lo":955,"Elo_hi":1047,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT","Elo":998,"Elo_lo":972,"Elo_hi":1024,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":986,"Elo_lo":939,"Elo_hi":1031,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":980,"Elo_lo":928,"Elo_hi":1034,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":976,"Elo_lo":910,"Elo_hi":1039,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":958,"Elo_lo":925,"Elo_hi":992,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":926,"Elo_lo":874,"Elo_hi":978,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":859,"Elo_lo":822,"Elo_hi":897,"n_targets":39,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":838,"Elo_lo":785,"Elo_hi":886,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":778,"Elo_lo":716,"Elo_hi":834,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n50","cell_label":"p=2,000, n=50","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":636,"Elo_lo":549,"Elo_hi":713,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1280,"Elo_lo":1214,"Elo_hi":1352,"n_targets":44,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1259,"Elo_lo":1195,"Elo_hi":1325,"n_targets":44,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1258,"Elo_lo":1184,"Elo_hi":1333,"n_targets":44,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1246,"Elo_lo":1169,"Elo_hi":1325,"n_targets":44,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1228,"Elo_lo":1157,"Elo_hi":1302,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1218,"Elo_lo":1173,"Elo_hi":1269,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1170,"Elo_lo":1097,"Elo_hi":1238,"n_targets":42,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1165,"Elo_lo":1111,"Elo_hi":1223,"n_targets":43,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1141,"Elo_lo":1077,"Elo_hi":1201,"n_targets":17,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1118,"Elo_lo":1046,"Elo_hi":1187,"n_targets":43,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1118,"Elo_lo":1065,"Elo_hi":1170,"n_targets":43,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1111,"Elo_lo":1052,"Elo_hi":1170,"n_targets":43,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1099,"Elo_lo":1042,"Elo_hi":1158,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1090,"Elo_lo":1036,"Elo_hi":1148,"n_targets":40,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1090,"Elo_lo":1033,"Elo_hi":1150,"n_targets":43,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1084,"Elo_lo":1017,"Elo_hi":1148,"n_targets":44,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1081,"Elo_lo":1022,"Elo_hi":1136,"n_targets":43,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1079,"Elo_lo":997,"Elo_hi":1155,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1065,"Elo_lo":982,"Elo_hi":1141,"n_targets":31,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1047,"Elo_lo":986,"Elo_hi":1109,"n_targets":40,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1044,"Elo_lo":993,"Elo_hi":1097,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1044,"Elo_lo":999,"Elo_hi":1095,"n_targets":44,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1042,"Elo_lo":962,"Elo_hi":1109,"n_targets":31,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1034,"Elo_lo":995,"Elo_hi":1073,"n_targets":40,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1021,"Elo_lo":997,"Elo_hi":1045,"n_targets":40,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1009,"Elo_lo":936,"Elo_hi":1075,"n_targets":44,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT","Elo":988,"Elo_lo":968,"Elo_hi":1007,"n_targets":44,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":985,"Elo_lo":951,"Elo_hi":1021,"n_targets":42,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":922,"Elo_lo":875,"Elo_hi":969,"n_targets":40,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":912,"Elo_lo":854,"Elo_hi":966,"n_targets":44,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":881,"Elo_lo":826,"Elo_hi":935,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":866,"Elo_lo":798,"Elo_hi":935,"n_targets":44,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n100","cell_label":"p=2,000, n=100","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":552,"Elo_lo":456,"Elo_hi":628,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1286,"Elo_lo":1211,"Elo_hi":1364,"n_targets":36,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1234,"Elo_lo":1157,"Elo_hi":1316,"n_targets":35,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1230,"Elo_lo":1157,"Elo_hi":1312,"n_targets":36,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1230,"Elo_lo":1141,"Elo_hi":1321,"n_targets":15,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1215,"Elo_lo":1136,"Elo_hi":1296,"n_targets":36,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1214,"Elo_lo":1153,"Elo_hi":1282,"n_targets":36,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1196,"Elo_lo":1127,"Elo_hi":1267,"n_targets":36,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1177,"Elo_lo":1113,"Elo_hi":1248,"n_targets":36,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1176,"Elo_lo":1101,"Elo_hi":1259,"n_targets":24,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1131,"Elo_lo":1047,"Elo_hi":1223,"n_targets":36,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1128,"Elo_lo":1062,"Elo_hi":1199,"n_targets":36,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1110,"Elo_lo":1039,"Elo_hi":1189,"n_targets":36,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1108,"Elo_lo":1032,"Elo_hi":1187,"n_targets":36,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1107,"Elo_lo":1020,"Elo_hi":1200,"n_targets":36,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1096,"Elo_lo":1024,"Elo_hi":1168,"n_targets":36,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1087,"Elo_lo":995,"Elo_hi":1179,"n_targets":24,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1078,"Elo_lo":1019,"Elo_hi":1138,"n_targets":34,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1078,"Elo_lo":988,"Elo_hi":1159,"n_targets":24,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1073,"Elo_lo":1009,"Elo_hi":1145,"n_targets":36,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1065,"Elo_lo":992,"Elo_hi":1135,"n_targets":36,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1060,"Elo_lo":984,"Elo_hi":1134,"n_targets":34,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1048,"Elo_lo":1003,"Elo_hi":1092,"n_targets":34,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1039,"Elo_lo":950,"Elo_hi":1124,"n_targets":36,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1016,"Elo_lo":948,"Elo_hi":1086,"n_targets":36,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1016,"Elo_lo":977,"Elo_hi":1054,"n_targets":34,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1012,"Elo_lo":952,"Elo_hi":1081,"n_targets":36,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":36,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":995,"Elo_lo":943,"Elo_hi":1048,"n_targets":35,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT","Elo":988,"Elo_lo":968,"Elo_hi":1008,"n_targets":36,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":955,"Elo_lo":899,"Elo_hi":1012,"n_targets":36,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":955,"Elo_lo":911,"Elo_hi":1002,"n_targets":34,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":910,"Elo_lo":853,"Elo_hi":970,"n_targets":36,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":831,"Elo_lo":754,"Elo_hi":905,"n_targets":36,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n200","cell_label":"p=2,000, n=200","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":452,"Elo_lo":351,"Elo_hi":542,"n_targets":36,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1236,"Elo_lo":1134,"Elo_hi":1355,"n_targets":22,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1222,"Elo_lo":1129,"Elo_hi":1330,"n_targets":21,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1215,"Elo_lo":1126,"Elo_hi":1313,"n_targets":22,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1184,"Elo_lo":1074,"Elo_hi":1309,"n_targets":22,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1158,"Elo_lo":1046,"Elo_hi":1275,"n_targets":22,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1144,"Elo_lo":1032,"Elo_hi":1268,"n_targets":22,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1135,"Elo_lo":1046,"Elo_hi":1232,"n_targets":22,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1127,"Elo_lo":1004,"Elo_hi":1251,"n_targets":7,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1113,"Elo_lo":1027,"Elo_hi":1210,"n_targets":22,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1098,"Elo_lo":990,"Elo_hi":1225,"n_targets":14,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1097,"Elo_lo":994,"Elo_hi":1218,"n_targets":14,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1071,"Elo_lo":1000,"Elo_hi":1147,"n_targets":22,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1057,"Elo_lo":954,"Elo_hi":1166,"n_targets":22,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1056,"Elo_lo":918,"Elo_hi":1211,"n_targets":21,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1042,"Elo_lo":964,"Elo_hi":1123,"n_targets":22,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1040,"Elo_lo":948,"Elo_hi":1128,"n_targets":22,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1027,"Elo_lo":941,"Elo_hi":1116,"n_targets":22,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1021,"Elo_lo":940,"Elo_hi":1107,"n_targets":22,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1017,"Elo_lo":917,"Elo_hi":1110,"n_targets":22,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1010,"Elo_lo":959,"Elo_hi":1060,"n_targets":22,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1009,"Elo_lo":866,"Elo_hi":1155,"n_targets":22,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1009,"Elo_lo":877,"Elo_hi":1147,"n_targets":14,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1006,"Elo_lo":977,"Elo_hi":1036,"n_targets":22,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT","Elo":1002,"Elo_lo":968,"Elo_hi":1036,"n_targets":22,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":22,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":986,"Elo_lo":902,"Elo_hi":1073,"n_targets":22,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":978,"Elo_lo":921,"Elo_hi":1035,"n_targets":22,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":977,"Elo_lo":918,"Elo_hi":1033,"n_targets":22,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":953,"Elo_lo":893,"Elo_hi":1022,"n_targets":22,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":914,"Elo_lo":814,"Elo_hi":1013,"n_targets":22,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR","Elo":900,"Elo_lo":761,"Elo_hi":1017,"n_targets":22,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":895,"Elo_lo":841,"Elo_hi":957,"n_targets":22,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":751,"Elo_lo":636,"Elo_hi":856,"n_targets":22,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n500","cell_label":"p=2,000, n=500","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":351,"Elo_lo":200,"Elo_hi":488,"n_targets":22,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1229,"Elo_lo":1117,"Elo_hi":1346,"n_targets":16,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1205,"Elo_lo":1067,"Elo_hi":1365,"n_targets":17,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1199,"Elo_lo":1090,"Elo_hi":1316,"n_targets":17,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1185,"Elo_lo":1059,"Elo_hi":1316,"n_targets":17,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1162,"Elo_lo":1025,"Elo_hi":1303,"n_targets":17,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1101,"Elo_lo":944,"Elo_hi":1269,"n_targets":17,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1093,"Elo_lo":943,"Elo_hi":1251,"n_targets":17,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1092,"Elo_lo":968,"Elo_hi":1211,"n_targets":10,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1084,"Elo_lo":937,"Elo_hi":1236,"n_targets":10,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1078,"Elo_lo":954,"Elo_hi":1228,"n_targets":17,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1038,"Elo_lo":909,"Elo_hi":1170,"n_targets":17,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1027,"Elo_lo":917,"Elo_hi":1145,"n_targets":17,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1011,"Elo_lo":928,"Elo_hi":1092,"n_targets":17,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1011,"Elo_lo":893,"Elo_hi":1119,"n_targets":2,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":17,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":997,"Elo_lo":858,"Elo_hi":1144,"n_targets":17,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":991,"Elo_lo":818,"Elo_hi":1186,"n_targets":16,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":990,"Elo_lo":918,"Elo_hi":1059,"n_targets":17,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT","Elo":989,"Elo_lo":957,"Elo_hi":1021,"n_targets":17,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":987,"Elo_lo":849,"Elo_hi":1115,"n_targets":17,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":975,"Elo_lo":837,"Elo_hi":1101,"n_targets":17,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":967,"Elo_lo":916,"Elo_hi":1018,"n_targets":17,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":955,"Elo_lo":880,"Elo_hi":1020,"n_targets":17,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":945,"Elo_lo":843,"Elo_hi":1040,"n_targets":17,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":929,"Elo_lo":835,"Elo_hi":1013,"n_targets":17,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":928,"Elo_lo":776,"Elo_hi":1069,"n_targets":17,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":924,"Elo_lo":718,"Elo_hi":1126,"n_targets":17,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":919,"Elo_lo":742,"Elo_hi":1103,"n_targets":10,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":851,"Elo_lo":699,"Elo_hi":984,"n_targets":17,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":835,"Elo_lo":708,"Elo_hi":977,"n_targets":17,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR","Elo":783,"Elo_lo":539,"Elo_hi":974,"n_targets":17,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":728,"Elo_lo":594,"Elo_hi":854,"n_targets":17,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":698,"Elo_lo":570,"Elo_hi":806,"n_targets":17,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000_n1000","cell_label":"p=2,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":317,"Elo_lo":135,"Elo_hi":478,"n_targets":17,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1274,"Elo_lo":1215,"Elo_hi":1337,"n_targets":45,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1259,"Elo_lo":1194,"Elo_hi":1326,"n_targets":45,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1250,"Elo_lo":1190,"Elo_hi":1313,"n_targets":46,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1238,"Elo_lo":1178,"Elo_hi":1300,"n_targets":45,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1231,"Elo_lo":1173,"Elo_hi":1297,"n_targets":40,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1231,"Elo_lo":1181,"Elo_hi":1283,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1198,"Elo_lo":1121,"Elo_hi":1276,"n_targets":32,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1181,"Elo_lo":1114,"Elo_hi":1254,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1180,"Elo_lo":1115,"Elo_hi":1252,"n_targets":45,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1174,"Elo_lo":1099,"Elo_hi":1251,"n_targets":5,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1151,"Elo_lo":1094,"Elo_hi":1215,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1133,"Elo_lo":1066,"Elo_hi":1206,"n_targets":31,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1128,"Elo_lo":1076,"Elo_hi":1185,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1120,"Elo_lo":1063,"Elo_hi":1180,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1098,"Elo_lo":1051,"Elo_hi":1148,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1093,"Elo_lo":1031,"Elo_hi":1157,"n_targets":44,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1084,"Elo_lo":1021,"Elo_hi":1146,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1078,"Elo_lo":1018,"Elo_hi":1136,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1077,"Elo_lo":1017,"Elo_hi":1137,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1059,"Elo_lo":981,"Elo_hi":1132,"n_targets":31,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1058,"Elo_lo":1019,"Elo_hi":1099,"n_targets":44,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1056,"Elo_lo":1007,"Elo_hi":1106,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1043,"Elo_lo":983,"Elo_hi":1103,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1031,"Elo_lo":1000,"Elo_hi":1060,"n_targets":45,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1014,"Elo_lo":965,"Elo_hi":1059,"n_targets":46,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1008,"Elo_lo":984,"Elo_hi":1032,"n_targets":45,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1002,"Elo_lo":920,"Elo_hi":1077,"n_targets":44,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR","Elo":997,"Elo_lo":920,"Elo_hi":1069,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT","Elo":992,"Elo_lo":968,"Elo_hi":1015,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":986,"Elo_lo":913,"Elo_hi":1053,"n_targets":44,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":986,"Elo_lo":941,"Elo_hi":1031,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":870,"Elo_lo":816,"Elo_hi":923,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_2000","cell_label":"p=2,000, n=all","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":363,"Elo_lo":251,"Elo_hi":451,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1149,"Elo_lo":1098,"Elo_hi":1204,"n_targets":45,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1143,"Elo_lo":1081,"Elo_hi":1208,"n_targets":45,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1130,"Elo_lo":1068,"Elo_hi":1194,"n_targets":44,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1100,"Elo_lo":1056,"Elo_hi":1145,"n_targets":45,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1090,"Elo_lo":1031,"Elo_hi":1148,"n_targets":44,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1073,"Elo_lo":1035,"Elo_hi":1111,"n_targets":44,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1072,"Elo_lo":1028,"Elo_hi":1119,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1069,"Elo_lo":1020,"Elo_hi":1117,"n_targets":45,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1058,"Elo_lo":1021,"Elo_hi":1095,"n_targets":44,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1056,"Elo_lo":1000,"Elo_hi":1111,"n_targets":45,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1019,"Elo_lo":961,"Elo_hi":1077,"n_targets":30,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1018,"Elo_lo":955,"Elo_hi":1082,"n_targets":44,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1015,"Elo_lo":958,"Elo_hi":1072,"n_targets":31,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1011,"Elo_lo":965,"Elo_hi":1057,"n_targets":44,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1007,"Elo_lo":875,"Elo_hi":1147,"n_targets":15,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT","Elo":1006,"Elo_lo":987,"Elo_hi":1025,"n_targets":45,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1004,"Elo_lo":970,"Elo_hi":1040,"n_targets":30,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1003,"Elo_lo":940,"Elo_hi":1066,"n_targets":43,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":45,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":995,"Elo_lo":965,"Elo_hi":1022,"n_targets":45,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":995,"Elo_lo":954,"Elo_hi":1039,"n_targets":30,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":978,"Elo_lo":943,"Elo_hi":1017,"n_targets":44,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":975,"Elo_lo":915,"Elo_hi":1034,"n_targets":31,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":973,"Elo_lo":914,"Elo_hi":1032,"n_targets":45,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":971,"Elo_lo":911,"Elo_hi":1034,"n_targets":30,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":965,"Elo_lo":929,"Elo_hi":1003,"n_targets":44,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":961,"Elo_lo":899,"Elo_hi":1020,"n_targets":45,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":944,"Elo_lo":895,"Elo_hi":994,"n_targets":45,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":904,"Elo_lo":857,"Elo_hi":950,"n_targets":45,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":899,"Elo_lo":848,"Elo_hi":953,"n_targets":45,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":852,"Elo_lo":795,"Elo_hi":903,"n_targets":30,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":840,"Elo_lo":796,"Elo_hi":883,"n_targets":45,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":737,"Elo_lo":658,"Elo_hi":807,"n_targets":45,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n20","cell_label":"p=10,000, n=20","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":732,"Elo_lo":654,"Elo_hi":805,"n_targets":45,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1258,"Elo_lo":1200,"Elo_hi":1319,"n_targets":33,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1230,"Elo_lo":1176,"Elo_hi":1287,"n_targets":45,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1226,"Elo_lo":1171,"Elo_hi":1279,"n_targets":46,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1221,"Elo_lo":1165,"Elo_hi":1280,"n_targets":46,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1187,"Elo_lo":1136,"Elo_hi":1242,"n_targets":45,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1146,"Elo_lo":1098,"Elo_hi":1193,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1141,"Elo_lo":1098,"Elo_hi":1185,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1098,"Elo_lo":1048,"Elo_hi":1146,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1090,"Elo_lo":1039,"Elo_hi":1146,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1081,"Elo_lo":1029,"Elo_hi":1126,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1073,"Elo_lo":1023,"Elo_hi":1122,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1073,"Elo_lo":974,"Elo_hi":1172,"n_targets":14,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1070,"Elo_lo":996,"Elo_hi":1150,"n_targets":44,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1068,"Elo_lo":1019,"Elo_hi":1119,"n_targets":39,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1056,"Elo_lo":997,"Elo_hi":1118,"n_targets":39,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1044,"Elo_lo":962,"Elo_hi":1120,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1037,"Elo_lo":987,"Elo_hi":1087,"n_targets":45,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1032,"Elo_lo":993,"Elo_hi":1072,"n_targets":39,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1025,"Elo_lo":972,"Elo_hi":1080,"n_targets":45,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1022,"Elo_lo":952,"Elo_hi":1092,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1013,"Elo_lo":987,"Elo_hi":1044,"n_targets":39,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1012,"Elo_lo":956,"Elo_hi":1069,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":999,"Elo_lo":944,"Elo_hi":1053,"n_targets":45,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":998,"Elo_lo":933,"Elo_hi":1063,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT","Elo":994,"Elo_lo":972,"Elo_hi":1017,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":966,"Elo_lo":923,"Elo_hi":1005,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":963,"Elo_lo":925,"Elo_hi":1003,"n_targets":46,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":962,"Elo_lo":912,"Elo_hi":1013,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":935,"Elo_lo":882,"Elo_hi":994,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":861,"Elo_lo":819,"Elo_hi":903,"n_targets":39,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":829,"Elo_lo":785,"Elo_hi":869,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":817,"Elo_lo":762,"Elo_hi":867,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n50","cell_label":"p=10,000, n=50","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":629,"Elo_lo":549,"Elo_hi":700,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1243,"Elo_lo":1184,"Elo_hi":1308,"n_targets":44,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1229,"Elo_lo":1160,"Elo_hi":1296,"n_targets":44,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1226,"Elo_lo":1162,"Elo_hi":1293,"n_targets":43,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1215,"Elo_lo":1154,"Elo_hi":1280,"n_targets":43,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1210,"Elo_lo":1151,"Elo_hi":1270,"n_targets":31,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1197,"Elo_lo":1154,"Elo_hi":1242,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1172,"Elo_lo":1127,"Elo_hi":1217,"n_targets":43,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1135,"Elo_lo":1059,"Elo_hi":1207,"n_targets":42,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1126,"Elo_lo":1059,"Elo_hi":1183,"n_targets":13,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1082,"Elo_lo":1037,"Elo_hi":1125,"n_targets":43,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1069,"Elo_lo":1013,"Elo_hi":1131,"n_targets":40,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1067,"Elo_lo":1005,"Elo_hi":1126,"n_targets":43,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1064,"Elo_lo":998,"Elo_hi":1128,"n_targets":44,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1056,"Elo_lo":1005,"Elo_hi":1106,"n_targets":43,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1052,"Elo_lo":974,"Elo_hi":1125,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1048,"Elo_lo":994,"Elo_hi":1106,"n_targets":40,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1044,"Elo_lo":971,"Elo_hi":1113,"n_targets":31,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"AUTOGLUON","Elo":1041,"Elo_lo":990,"Elo_hi":1091,"n_targets":44,"id":"AUTOGLUON","display":"AutoGluon","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"AutoGluon automates tabular learning by training multiple models and combining their predictions in ensembles. Here it runs its native extreme preset within a one-hour budget as a separate AutoML reference.","tuning":null,"category":"AutoML","color":"#111827","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1041,"Elo_lo":984,"Elo_hi":1098,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1041,"Elo_lo":988,"Elo_hi":1094,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1033,"Elo_lo":991,"Elo_hi":1072,"n_targets":40,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1029,"Elo_lo":965,"Elo_hi":1095,"n_targets":43,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1029,"Elo_lo":946,"Elo_hi":1099,"n_targets":31,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1011,"Elo_lo":969,"Elo_hi":1064,"n_targets":44,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1007,"Elo_lo":937,"Elo_hi":1073,"n_targets":44,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1005,"Elo_lo":977,"Elo_hi":1037,"n_targets":40,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":994,"Elo_lo":957,"Elo_hi":1035,"n_targets":42,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT","Elo":977,"Elo_lo":951,"Elo_hi":999,"n_targets":44,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":973,"Elo_lo":924,"Elo_hi":1023,"n_targets":43,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":918,"Elo_lo":869,"Elo_hi":967,"n_targets":40,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":910,"Elo_lo":853,"Elo_hi":964,"n_targets":44,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":875,"Elo_lo":823,"Elo_hi":925,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":868,"Elo_lo":813,"Elo_hi":923,"n_targets":44,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n100","cell_label":"p=10,000, n=100","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":526,"Elo_lo":434,"Elo_hi":601,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1263,"Elo_lo":1200,"Elo_hi":1332,"n_targets":36,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1223,"Elo_lo":1170,"Elo_hi":1281,"n_targets":36,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1221,"Elo_lo":1138,"Elo_hi":1300,"n_targets":13,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1213,"Elo_lo":1137,"Elo_hi":1292,"n_targets":36,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1212,"Elo_lo":1148,"Elo_hi":1282,"n_targets":36,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1206,"Elo_lo":1136,"Elo_hi":1272,"n_targets":36,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1175,"Elo_lo":1095,"Elo_hi":1261,"n_targets":35,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1164,"Elo_lo":1110,"Elo_hi":1228,"n_targets":36,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1161,"Elo_lo":1091,"Elo_hi":1241,"n_targets":24,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1122,"Elo_lo":1062,"Elo_hi":1188,"n_targets":36,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1100,"Elo_lo":1010,"Elo_hi":1189,"n_targets":24,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1099,"Elo_lo":1040,"Elo_hi":1156,"n_targets":34,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1080,"Elo_lo":997,"Elo_hi":1161,"n_targets":36,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1079,"Elo_lo":1009,"Elo_hi":1154,"n_targets":36,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1077,"Elo_lo":999,"Elo_hi":1162,"n_targets":36,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1068,"Elo_lo":1003,"Elo_hi":1131,"n_targets":36,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1064,"Elo_lo":999,"Elo_hi":1133,"n_targets":36,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1061,"Elo_lo":995,"Elo_hi":1125,"n_targets":36,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1057,"Elo_lo":978,"Elo_hi":1133,"n_targets":36,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1053,"Elo_lo":961,"Elo_hi":1147,"n_targets":24,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1043,"Elo_lo":1011,"Elo_hi":1078,"n_targets":34,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1026,"Elo_lo":985,"Elo_hi":1073,"n_targets":34,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1022,"Elo_lo":957,"Elo_hi":1082,"n_targets":34,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1020,"Elo_lo":979,"Elo_hi":1066,"n_targets":35,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":36,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":999,"Elo_lo":941,"Elo_hi":1065,"n_targets":36,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":993,"Elo_lo":923,"Elo_hi":1064,"n_targets":36,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":993,"Elo_lo":940,"Elo_hi":1053,"n_targets":34,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":982,"Elo_lo":928,"Elo_hi":1039,"n_targets":36,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT","Elo":982,"Elo_lo":957,"Elo_hi":1002,"n_targets":36,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":966,"Elo_lo":898,"Elo_hi":1033,"n_targets":36,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":926,"Elo_lo":864,"Elo_hi":995,"n_targets":36,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":837,"Elo_lo":773,"Elo_hi":897,"n_targets":36,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n200","cell_label":"p=10,000, n=200","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":457,"Elo_lo":356,"Elo_hi":543,"n_targets":36,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1217,"Elo_lo":1118,"Elo_hi":1331,"n_targets":22,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1187,"Elo_lo":1116,"Elo_hi":1274,"n_targets":22,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1185,"Elo_lo":1095,"Elo_hi":1290,"n_targets":20,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1162,"Elo_lo":1053,"Elo_hi":1275,"n_targets":22,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1161,"Elo_lo":1055,"Elo_hi":1268,"n_targets":22,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1149,"Elo_lo":1029,"Elo_hi":1248,"n_targets":6,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1137,"Elo_lo":1010,"Elo_hi":1270,"n_targets":14,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1117,"Elo_lo":1028,"Elo_hi":1210,"n_targets":22,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1107,"Elo_lo":1002,"Elo_hi":1221,"n_targets":22,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1091,"Elo_lo":1013,"Elo_hi":1179,"n_targets":22,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1086,"Elo_lo":974,"Elo_hi":1207,"n_targets":14,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1062,"Elo_lo":984,"Elo_hi":1143,"n_targets":22,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1049,"Elo_lo":953,"Elo_hi":1139,"n_targets":22,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1046,"Elo_lo":943,"Elo_hi":1154,"n_targets":22,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1046,"Elo_lo":965,"Elo_hi":1130,"n_targets":22,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1032,"Elo_lo":968,"Elo_hi":1088,"n_targets":22,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1030,"Elo_lo":995,"Elo_hi":1069,"n_targets":22,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1025,"Elo_lo":929,"Elo_hi":1114,"n_targets":22,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1015,"Elo_lo":921,"Elo_hi":1106,"n_targets":22,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1014,"Elo_lo":931,"Elo_hi":1093,"n_targets":22,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1005,"Elo_lo":859,"Elo_hi":1145,"n_targets":22,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":22,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT","Elo":999,"Elo_lo":971,"Elo_hi":1024,"n_targets":22,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":994,"Elo_lo":917,"Elo_hi":1067,"n_targets":22,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":983,"Elo_lo":852,"Elo_hi":1116,"n_targets":14,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":977,"Elo_lo":887,"Elo_hi":1064,"n_targets":22,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":976,"Elo_lo":918,"Elo_hi":1038,"n_targets":22,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":954,"Elo_lo":812,"Elo_hi":1132,"n_targets":13,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR","Elo":944,"Elo_lo":794,"Elo_hi":1070,"n_targets":22,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":936,"Elo_lo":834,"Elo_hi":1032,"n_targets":22,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":925,"Elo_lo":852,"Elo_hi":992,"n_targets":22,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":881,"Elo_lo":785,"Elo_hi":971,"n_targets":22,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":744,"Elo_lo":626,"Elo_hi":840,"n_targets":22,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n500","cell_label":"p=10,000, n=500","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":355,"Elo_lo":201,"Elo_hi":498,"n_targets":22,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1255,"Elo_lo":1141,"Elo_hi":1384,"n_targets":13,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1206,"Elo_lo":1076,"Elo_hi":1355,"n_targets":17,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1197,"Elo_lo":1120,"Elo_hi":1302,"n_targets":17,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1171,"Elo_lo":1054,"Elo_hi":1289,"n_targets":17,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1130,"Elo_lo":992,"Elo_hi":1281,"n_targets":17,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1110,"Elo_lo":987,"Elo_hi":1231,"n_targets":10,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1110,"Elo_lo":969,"Elo_hi":1270,"n_targets":17,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1082,"Elo_lo":967,"Elo_hi":1220,"n_targets":17,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1077,"Elo_lo":954,"Elo_hi":1222,"n_targets":17,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1060,"Elo_lo":903,"Elo_hi":1225,"n_targets":10,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1043,"Elo_lo":937,"Elo_hi":1156,"n_targets":17,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1023,"Elo_lo":905,"Elo_hi":1144,"n_targets":17,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1011,"Elo_lo":899,"Elo_hi":1118,"n_targets":2,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT","Elo":1003,"Elo_lo":969,"Elo_hi":1041,"n_targets":17,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":17,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":995,"Elo_lo":910,"Elo_hi":1085,"n_targets":17,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":993,"Elo_lo":952,"Elo_hi":1032,"n_targets":17,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":990,"Elo_lo":914,"Elo_hi":1069,"n_targets":17,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":989,"Elo_lo":861,"Elo_hi":1123,"n_targets":17,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":987,"Elo_lo":843,"Elo_hi":1139,"n_targets":17,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":986,"Elo_lo":931,"Elo_hi":1037,"n_targets":17,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":979,"Elo_lo":895,"Elo_hi":1056,"n_targets":17,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":974,"Elo_lo":839,"Elo_hi":1115,"n_targets":17,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":961,"Elo_lo":818,"Elo_hi":1104,"n_targets":17,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":946,"Elo_lo":793,"Elo_hi":1136,"n_targets":13,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":931,"Elo_lo":847,"Elo_hi":1015,"n_targets":17,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":920,"Elo_lo":721,"Elo_hi":1115,"n_targets":17,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":891,"Elo_lo":700,"Elo_hi":1087,"n_targets":10,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":860,"Elo_lo":739,"Elo_hi":991,"n_targets":17,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":837,"Elo_lo":683,"Elo_hi":974,"n_targets":17,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR","Elo":804,"Elo_lo":561,"Elo_hi":1003,"n_targets":17,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":697,"Elo_lo":561,"Elo_hi":824,"n_targets":16,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":689,"Elo_lo":568,"Elo_hi":799,"n_targets":17,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000_n1000","cell_label":"p=10,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":319,"Elo_lo":151,"Elo_hi":474,"n_targets":17,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1265,"Elo_lo":1204,"Elo_hi":1325,"n_targets":44,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1256,"Elo_lo":1188,"Elo_hi":1325,"n_targets":45,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1220,"Elo_lo":1158,"Elo_hi":1282,"n_targets":45,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1217,"Elo_lo":1167,"Elo_hi":1268,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1214,"Elo_lo":1151,"Elo_hi":1279,"n_targets":45,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1191,"Elo_lo":1115,"Elo_hi":1267,"n_targets":29,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1185,"Elo_lo":1121,"Elo_hi":1253,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1184,"Elo_lo":1122,"Elo_hi":1252,"n_targets":34,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1183,"Elo_lo":1086,"Elo_hi":1270,"n_targets":4,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1139,"Elo_lo":1079,"Elo_hi":1205,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1131,"Elo_lo":1055,"Elo_hi":1203,"n_targets":24,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1122,"Elo_lo":1063,"Elo_hi":1180,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1122,"Elo_lo":1065,"Elo_hi":1183,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1119,"Elo_lo":1050,"Elo_hi":1192,"n_targets":31,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1095,"Elo_lo":1031,"Elo_hi":1153,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1086,"Elo_lo":1041,"Elo_hi":1134,"n_targets":44,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1082,"Elo_lo":1028,"Elo_hi":1135,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1081,"Elo_lo":1017,"Elo_hi":1145,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1068,"Elo_lo":1009,"Elo_hi":1125,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1064,"Elo_lo":996,"Elo_hi":1129,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1059,"Elo_lo":1011,"Elo_hi":1105,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1049,"Elo_lo":982,"Elo_hi":1115,"n_targets":44,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1046,"Elo_lo":965,"Elo_hi":1124,"n_targets":31,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1044,"Elo_lo":1017,"Elo_hi":1073,"n_targets":46,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1034,"Elo_lo":989,"Elo_hi":1078,"n_targets":45,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1025,"Elo_lo":996,"Elo_hi":1056,"n_targets":45,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1015,"Elo_lo":938,"Elo_hi":1088,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1005,"Elo_lo":932,"Elo_hi":1074,"n_targets":44,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":1001,"Elo_lo":949,"Elo_hi":1052,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT","Elo":983,"Elo_lo":958,"Elo_hi":1009,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":877,"Elo_lo":828,"Elo_hi":928,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":802,"Elo_lo":715,"Elo_hi":875,"n_targets":42,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_10000","cell_label":"p=10,000, n=all","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":342,"Elo_lo":229,"Elo_hi":437,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1140,"Elo_lo":1081,"Elo_hi":1201,"n_targets":44,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1123,"Elo_lo":1059,"Elo_hi":1192,"n_targets":44,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1085,"Elo_lo":1015,"Elo_hi":1163,"n_targets":34,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1081,"Elo_lo":1034,"Elo_hi":1133,"n_targets":45,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1062,"Elo_lo":1020,"Elo_hi":1103,"n_targets":44,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1060,"Elo_lo":1013,"Elo_hi":1111,"n_targets":31,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1052,"Elo_lo":1010,"Elo_hi":1094,"n_targets":44,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1040,"Elo_lo":990,"Elo_hi":1091,"n_targets":45,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1037,"Elo_lo":981,"Elo_hi":1094,"n_targets":44,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1036,"Elo_lo":980,"Elo_hi":1089,"n_targets":45,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1009,"Elo_lo":944,"Elo_hi":1074,"n_targets":44,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":45,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":996,"Elo_lo":937,"Elo_hi":1051,"n_targets":30,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":992,"Elo_lo":941,"Elo_hi":1043,"n_targets":44,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT","Elo":992,"Elo_lo":973,"Elo_hi":1009,"n_targets":45,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":991,"Elo_lo":935,"Elo_hi":1051,"n_targets":31,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":989,"Elo_lo":958,"Elo_hi":1023,"n_targets":30,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":989,"Elo_lo":852,"Elo_hi":1128,"n_targets":15,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":975,"Elo_lo":908,"Elo_hi":1036,"n_targets":45,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":973,"Elo_lo":913,"Elo_hi":1032,"n_targets":43,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":973,"Elo_lo":934,"Elo_hi":1014,"n_targets":30,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":971,"Elo_lo":911,"Elo_hi":1031,"n_targets":45,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":967,"Elo_lo":931,"Elo_hi":998,"n_targets":45,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":958,"Elo_lo":919,"Elo_hi":999,"n_targets":44,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":950,"Elo_lo":886,"Elo_hi":1014,"n_targets":31,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":936,"Elo_lo":894,"Elo_hi":979,"n_targets":43,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":935,"Elo_lo":871,"Elo_hi":998,"n_targets":30,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":930,"Elo_lo":877,"Elo_hi":982,"n_targets":45,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":885,"Elo_lo":838,"Elo_hi":931,"n_targets":45,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":879,"Elo_lo":823,"Elo_hi":936,"n_targets":45,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":831,"Elo_lo":767,"Elo_hi":890,"n_targets":30,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":809,"Elo_lo":760,"Elo_hi":856,"n_targets":45,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":726,"Elo_lo":643,"Elo_hi":799,"n_targets":45,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n20","cell_label":"p=25,000, n=20","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":722,"Elo_lo":640,"Elo_hi":799,"n_targets":45,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1246,"Elo_lo":1191,"Elo_hi":1305,"n_targets":32,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1230,"Elo_lo":1181,"Elo_hi":1282,"n_targets":45,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1222,"Elo_lo":1167,"Elo_hi":1276,"n_targets":45,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1194,"Elo_lo":1143,"Elo_hi":1252,"n_targets":33,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1154,"Elo_lo":1114,"Elo_hi":1195,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1154,"Elo_lo":1104,"Elo_hi":1205,"n_targets":45,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1147,"Elo_lo":1100,"Elo_hi":1192,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1097,"Elo_lo":1050,"Elo_hi":1143,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1077,"Elo_lo":1028,"Elo_hi":1129,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1076,"Elo_lo":1023,"Elo_hi":1120,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1071,"Elo_lo":1019,"Elo_hi":1124,"n_targets":45,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1065,"Elo_lo":1012,"Elo_hi":1118,"n_targets":39,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1064,"Elo_lo":971,"Elo_hi":1161,"n_targets":14,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1051,"Elo_lo":978,"Elo_hi":1125,"n_targets":44,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1039,"Elo_lo":982,"Elo_hi":1099,"n_targets":39,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1034,"Elo_lo":956,"Elo_hi":1108,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1026,"Elo_lo":989,"Elo_hi":1065,"n_targets":39,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1019,"Elo_lo":966,"Elo_hi":1072,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1004,"Elo_lo":948,"Elo_hi":1061,"n_targets":44,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1004,"Elo_lo":937,"Elo_hi":1068,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1001,"Elo_lo":973,"Elo_hi":1031,"n_targets":39,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"XT","Elo":995,"Elo_lo":974,"Elo_hi":1014,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":993,"Elo_lo":932,"Elo_hi":1051,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":973,"Elo_lo":934,"Elo_hi":1011,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":971,"Elo_lo":916,"Elo_hi":1021,"n_targets":41,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":962,"Elo_lo":927,"Elo_hi":1003,"n_targets":45,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":958,"Elo_lo":911,"Elo_hi":1008,"n_targets":45,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":948,"Elo_lo":895,"Elo_hi":1005,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":937,"Elo_lo":886,"Elo_hi":992,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":867,"Elo_lo":819,"Elo_hi":918,"n_targets":39,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":843,"Elo_lo":793,"Elo_hi":889,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":828,"Elo_lo":775,"Elo_hi":876,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n50","cell_label":"p=25,000, n=50","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":630,"Elo_lo":551,"Elo_hi":701,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1259,"Elo_lo":1197,"Elo_hi":1327,"n_targets":43,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1241,"Elo_lo":1171,"Elo_hi":1310,"n_targets":43,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1240,"Elo_lo":1181,"Elo_hi":1304,"n_targets":30,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1221,"Elo_lo":1148,"Elo_hi":1297,"n_targets":29,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1213,"Elo_lo":1155,"Elo_hi":1275,"n_targets":43,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1199,"Elo_lo":1157,"Elo_hi":1244,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1191,"Elo_lo":1141,"Elo_hi":1244,"n_targets":43,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1130,"Elo_lo":1059,"Elo_hi":1187,"n_targets":13,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1128,"Elo_lo":1058,"Elo_hi":1198,"n_targets":42,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1096,"Elo_lo":1050,"Elo_hi":1140,"n_targets":43,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1085,"Elo_lo":1020,"Elo_hi":1146,"n_targets":43,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1070,"Elo_lo":1011,"Elo_hi":1128,"n_targets":40,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1063,"Elo_lo":1002,"Elo_hi":1125,"n_targets":44,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1056,"Elo_lo":983,"Elo_hi":1127,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1051,"Elo_lo":994,"Elo_hi":1109,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1051,"Elo_lo":1004,"Elo_hi":1097,"n_targets":40,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1045,"Elo_lo":970,"Elo_hi":1114,"n_targets":31,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1044,"Elo_lo":978,"Elo_hi":1107,"n_targets":40,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1040,"Elo_lo":984,"Elo_hi":1095,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1039,"Elo_lo":954,"Elo_hi":1115,"n_targets":31,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1032,"Elo_lo":966,"Elo_hi":1096,"n_targets":40,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1028,"Elo_lo":959,"Elo_hi":1094,"n_targets":44,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1027,"Elo_lo":970,"Elo_hi":1083,"n_targets":39,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1016,"Elo_lo":988,"Elo_hi":1051,"n_targets":40,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1016,"Elo_lo":967,"Elo_hi":1066,"n_targets":44,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1006,"Elo_lo":970,"Elo_hi":1044,"n_targets":41,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"XT","Elo":988,"Elo_lo":970,"Elo_hi":1004,"n_targets":44,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":933,"Elo_lo":873,"Elo_hi":990,"n_targets":43,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":923,"Elo_lo":865,"Elo_hi":973,"n_targets":44,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":923,"Elo_lo":876,"Elo_hi":969,"n_targets":40,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":880,"Elo_lo":829,"Elo_hi":923,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":879,"Elo_lo":817,"Elo_hi":939,"n_targets":44,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n100","cell_label":"p=25,000, n=100","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":546,"Elo_lo":454,"Elo_hi":616,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1265,"Elo_lo":1203,"Elo_hi":1336,"n_targets":36,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1218,"Elo_lo":1139,"Elo_hi":1296,"n_targets":13,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1208,"Elo_lo":1139,"Elo_hi":1276,"n_targets":36,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1205,"Elo_lo":1143,"Elo_hi":1272,"n_targets":36,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1201,"Elo_lo":1149,"Elo_hi":1261,"n_targets":36,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1183,"Elo_lo":1099,"Elo_hi":1266,"n_targets":26,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1176,"Elo_lo":1094,"Elo_hi":1261,"n_targets":32,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1174,"Elo_lo":1119,"Elo_hi":1233,"n_targets":36,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1160,"Elo_lo":1095,"Elo_hi":1229,"n_targets":24,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1114,"Elo_lo":1058,"Elo_hi":1176,"n_targets":36,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1102,"Elo_lo":1046,"Elo_hi":1155,"n_targets":34,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1079,"Elo_lo":992,"Elo_hi":1176,"n_targets":26,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1078,"Elo_lo":996,"Elo_hi":1161,"n_targets":36,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1075,"Elo_lo":990,"Elo_hi":1159,"n_targets":24,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1067,"Elo_lo":1005,"Elo_hi":1129,"n_targets":36,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1062,"Elo_lo":1002,"Elo_hi":1127,"n_targets":36,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1056,"Elo_lo":980,"Elo_hi":1131,"n_targets":36,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1056,"Elo_lo":997,"Elo_hi":1114,"n_targets":36,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1054,"Elo_lo":984,"Elo_hi":1124,"n_targets":34,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1049,"Elo_lo":1008,"Elo_hi":1095,"n_targets":34,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1044,"Elo_lo":955,"Elo_hi":1129,"n_targets":24,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1028,"Elo_lo":967,"Elo_hi":1087,"n_targets":34,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1028,"Elo_lo":992,"Elo_hi":1069,"n_targets":34,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1023,"Elo_lo":957,"Elo_hi":1084,"n_targets":36,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1016,"Elo_lo":979,"Elo_hi":1054,"n_targets":33,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":36,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":997,"Elo_lo":941,"Elo_hi":1063,"n_targets":36,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":993,"Elo_lo":936,"Elo_hi":1048,"n_targets":36,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":990,"Elo_lo":941,"Elo_hi":1046,"n_targets":33,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"XT","Elo":990,"Elo_lo":970,"Elo_hi":1009,"n_targets":36,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":975,"Elo_lo":906,"Elo_hi":1051,"n_targets":30,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":927,"Elo_lo":867,"Elo_hi":993,"n_targets":36,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":842,"Elo_lo":770,"Elo_hi":909,"n_targets":36,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n200","cell_label":"p=25,000, n=200","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":467,"Elo_lo":372,"Elo_hi":549,"n_targets":36,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1214,"Elo_lo":1114,"Elo_hi":1334,"n_targets":22,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1196,"Elo_lo":1098,"Elo_hi":1313,"n_targets":14,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1178,"Elo_lo":1102,"Elo_hi":1273,"n_targets":22,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1150,"Elo_lo":1026,"Elo_hi":1247,"n_targets":6,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1142,"Elo_lo":1038,"Elo_hi":1251,"n_targets":22,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1134,"Elo_lo":1009,"Elo_hi":1263,"n_targets":14,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1130,"Elo_lo":1028,"Elo_hi":1240,"n_targets":18,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1125,"Elo_lo":1034,"Elo_hi":1221,"n_targets":22,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1108,"Elo_lo":1001,"Elo_hi":1214,"n_targets":22,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1080,"Elo_lo":999,"Elo_hi":1173,"n_targets":22,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1070,"Elo_lo":964,"Elo_hi":1179,"n_targets":14,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1068,"Elo_lo":989,"Elo_hi":1151,"n_targets":22,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1053,"Elo_lo":955,"Elo_hi":1149,"n_targets":22,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1049,"Elo_lo":968,"Elo_hi":1132,"n_targets":22,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1040,"Elo_lo":944,"Elo_hi":1140,"n_targets":21,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1029,"Elo_lo":996,"Elo_hi":1063,"n_targets":22,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1021,"Elo_lo":924,"Elo_hi":1113,"n_targets":22,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1017,"Elo_lo":956,"Elo_hi":1074,"n_targets":22,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1006,"Elo_lo":861,"Elo_hi":1147,"n_targets":22,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":22,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"XT","Elo":999,"Elo_lo":971,"Elo_hi":1023,"n_targets":22,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":996,"Elo_lo":899,"Elo_hi":1094,"n_targets":22,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":995,"Elo_lo":914,"Elo_hi":1072,"n_targets":22,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":993,"Elo_lo":900,"Elo_hi":1082,"n_targets":22,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":984,"Elo_lo":912,"Elo_hi":1059,"n_targets":22,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":975,"Elo_lo":921,"Elo_hi":1036,"n_targets":21,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":969,"Elo_lo":825,"Elo_hi":1111,"n_targets":14,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":955,"Elo_lo":811,"Elo_hi":1131,"n_targets":13,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":935,"Elo_lo":853,"Elo_hi":1013,"n_targets":22,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"LR","Elo":932,"Elo_lo":786,"Elo_hi":1061,"n_targets":22,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":928,"Elo_lo":826,"Elo_hi":1024,"n_targets":22,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":903,"Elo_lo":803,"Elo_hi":1001,"n_targets":18,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":732,"Elo_lo":614,"Elo_hi":834,"n_targets":22,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n500","cell_label":"p=25,000, n=500","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":363,"Elo_lo":219,"Elo_hi":506,"n_targets":22,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1253,"Elo_lo":1139,"Elo_hi":1384,"n_targets":13,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1210,"Elo_lo":1079,"Elo_hi":1362,"n_targets":17,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1210,"Elo_lo":1127,"Elo_hi":1318,"n_targets":17,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1162,"Elo_lo":1049,"Elo_hi":1285,"n_targets":17,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1114,"Elo_lo":971,"Elo_hi":1277,"n_targets":17,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1113,"Elo_lo":989,"Elo_hi":1241,"n_targets":10,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1113,"Elo_lo":985,"Elo_hi":1252,"n_targets":17,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1085,"Elo_lo":969,"Elo_hi":1224,"n_targets":17,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1084,"Elo_lo":956,"Elo_hi":1232,"n_targets":17,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1059,"Elo_lo":899,"Elo_hi":1233,"n_targets":10,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1052,"Elo_lo":941,"Elo_hi":1169,"n_targets":17,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1015,"Elo_lo":900,"Elo_hi":1133,"n_targets":17,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1010,"Elo_lo":899,"Elo_hi":1117,"n_targets":2,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1001,"Elo_lo":912,"Elo_hi":1091,"n_targets":17,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":17,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT","Elo":997,"Elo_lo":961,"Elo_hi":1036,"n_targets":17,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":993,"Elo_lo":951,"Elo_hi":1032,"n_targets":17,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":984,"Elo_lo":840,"Elo_hi":1135,"n_targets":17,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":983,"Elo_lo":896,"Elo_hi":1060,"n_targets":17,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":979,"Elo_lo":852,"Elo_hi":1116,"n_targets":17,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":978,"Elo_lo":907,"Elo_hi":1048,"n_targets":17,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":978,"Elo_lo":928,"Elo_hi":1026,"n_targets":17,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":968,"Elo_lo":823,"Elo_hi":1115,"n_targets":17,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":956,"Elo_lo":831,"Elo_hi":1087,"n_targets":17,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":945,"Elo_lo":792,"Elo_hi":1136,"n_targets":13,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":927,"Elo_lo":729,"Elo_hi":1129,"n_targets":17,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":923,"Elo_lo":846,"Elo_hi":996,"n_targets":17,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":901,"Elo_lo":700,"Elo_hi":1106,"n_targets":10,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":860,"Elo_lo":739,"Elo_hi":996,"n_targets":17,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":831,"Elo_lo":678,"Elo_hi":964,"n_targets":17,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR","Elo":794,"Elo_lo":557,"Elo_hi":988,"n_targets":17,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":718,"Elo_lo":587,"Elo_hi":841,"n_targets":15,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":692,"Elo_lo":570,"Elo_hi":803,"n_targets":17,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000_n1000","cell_label":"p=25,000, n=1,000","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":320,"Elo_lo":151,"Elo_hi":474,"n_targets":17,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1246,"Elo_lo":1185,"Elo_hi":1317,"n_targets":30,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1227,"Elo_lo":1161,"Elo_hi":1294,"n_targets":44,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1217,"Elo_lo":1155,"Elo_hi":1278,"n_targets":44,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1211,"Elo_lo":1159,"Elo_hi":1265,"n_targets":46,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1199,"Elo_lo":1123,"Elo_hi":1277,"n_targets":28,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1195,"Elo_lo":1136,"Elo_hi":1256,"n_targets":45,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1171,"Elo_lo":1078,"Elo_hi":1249,"n_targets":4,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1165,"Elo_lo":1101,"Elo_hi":1233,"n_targets":46,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1162,"Elo_lo":1097,"Elo_hi":1236,"n_targets":27,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1143,"Elo_lo":1083,"Elo_hi":1203,"n_targets":45,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1106,"Elo_lo":1050,"Elo_hi":1170,"n_targets":45,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1102,"Elo_lo":1046,"Elo_hi":1160,"n_targets":46,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1096,"Elo_lo":1005,"Elo_hi":1180,"n_targets":16,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1093,"Elo_lo":1024,"Elo_hi":1164,"n_targets":31,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1089,"Elo_lo":1046,"Elo_hi":1134,"n_targets":43,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1083,"Elo_lo":1019,"Elo_hi":1145,"n_targets":46,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1072,"Elo_lo":1023,"Elo_hi":1120,"n_targets":46,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1071,"Elo_lo":1011,"Elo_hi":1126,"n_targets":46,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1053,"Elo_lo":1000,"Elo_hi":1105,"n_targets":46,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1050,"Elo_lo":999,"Elo_hi":1101,"n_targets":46,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1046,"Elo_lo":981,"Elo_hi":1109,"n_targets":46,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1041,"Elo_lo":998,"Elo_hi":1087,"n_targets":44,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1035,"Elo_lo":998,"Elo_hi":1071,"n_targets":45,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1032,"Elo_lo":951,"Elo_hi":1107,"n_targets":31,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1027,"Elo_lo":960,"Elo_hi":1092,"n_targets":42,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1022,"Elo_lo":996,"Elo_hi":1052,"n_targets":45,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1015,"Elo_lo":940,"Elo_hi":1089,"n_targets":46,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":46,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":990,"Elo_lo":917,"Elo_hi":1061,"n_targets":44,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"XT","Elo":986,"Elo_lo":962,"Elo_hi":1009,"n_targets":46,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":980,"Elo_lo":934,"Elo_hi":1027,"n_targets":46,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":862,"Elo_lo":811,"Elo_hi":914,"n_targets":46,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":808,"Elo_lo":716,"Elo_hi":886,"n_targets":29,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_25000","cell_label":"p=25,000, n=all","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":360,"Elo_lo":255,"Elo_hi":442,"n_targets":46,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1141,"Elo_lo":1084,"Elo_hi":1200,"n_targets":42,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1118,"Elo_lo":1055,"Elo_hi":1186,"n_targets":42,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1087,"Elo_lo":1015,"Elo_hi":1163,"n_targets":32,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1087,"Elo_lo":1040,"Elo_hi":1136,"n_targets":43,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1075,"Elo_lo":1028,"Elo_hi":1121,"n_targets":34,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1067,"Elo_lo":1025,"Elo_hi":1111,"n_targets":37,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1058,"Elo_lo":1008,"Elo_hi":1107,"n_targets":29,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1043,"Elo_lo":993,"Elo_hi":1094,"n_targets":43,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1040,"Elo_lo":983,"Elo_hi":1090,"n_targets":43,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1033,"Elo_lo":973,"Elo_hi":1092,"n_targets":41,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1005,"Elo_lo":938,"Elo_hi":1068,"n_targets":42,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":43,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":997,"Elo_lo":943,"Elo_hi":1049,"n_targets":28,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"XT","Elo":997,"Elo_lo":976,"Elo_hi":1017,"n_targets":42,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":996,"Elo_lo":945,"Elo_hi":1051,"n_targets":38,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":992,"Elo_lo":958,"Elo_hi":1026,"n_targets":28,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":992,"Elo_lo":852,"Elo_hi":1126,"n_targets":15,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":990,"Elo_lo":928,"Elo_hi":1049,"n_targets":37,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":984,"Elo_lo":925,"Elo_hi":1046,"n_targets":31,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":983,"Elo_lo":945,"Elo_hi":1021,"n_targets":28,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":981,"Elo_lo":917,"Elo_hi":1042,"n_targets":43,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":973,"Elo_lo":928,"Elo_hi":1015,"n_targets":40,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":969,"Elo_lo":932,"Elo_hi":1004,"n_targets":39,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":968,"Elo_lo":910,"Elo_hi":1030,"n_targets":43,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":940,"Elo_lo":893,"Elo_hi":984,"n_targets":40,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":938,"Elo_lo":863,"Elo_hi":1005,"n_targets":28,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":936,"Elo_lo":871,"Elo_hi":999,"n_targets":31,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":928,"Elo_lo":870,"Elo_hi":978,"n_targets":42,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":891,"Elo_lo":845,"Elo_hi":932,"n_targets":43,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":884,"Elo_lo":826,"Elo_hi":940,"n_targets":43,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":834,"Elo_lo":765,"Elo_hi":898,"n_targets":28,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":821,"Elo_lo":771,"Elo_hi":870,"n_targets":43,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":725,"Elo_lo":639,"Elo_hi":801,"n_targets":43,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n20","cell_label":"p=all, n=20","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":721,"Elo_lo":637,"Elo_hi":795,"n_targets":43,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1237,"Elo_lo":1182,"Elo_hi":1295,"n_targets":30,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1222,"Elo_lo":1170,"Elo_hi":1272,"n_targets":43,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1216,"Elo_lo":1161,"Elo_hi":1277,"n_targets":43,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1197,"Elo_lo":1141,"Elo_hi":1261,"n_targets":31,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1156,"Elo_lo":1104,"Elo_hi":1210,"n_targets":42,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1154,"Elo_lo":1114,"Elo_hi":1192,"n_targets":37,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1137,"Elo_lo":1090,"Elo_hi":1184,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1100,"Elo_lo":1052,"Elo_hi":1145,"n_targets":34,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1087,"Elo_lo":1011,"Elo_hi":1176,"n_targets":37,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1085,"Elo_lo":1030,"Elo_hi":1140,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1066,"Elo_lo":1015,"Elo_hi":1115,"n_targets":43,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1065,"Elo_lo":967,"Elo_hi":1161,"n_targets":14,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1062,"Elo_lo":1012,"Elo_hi":1116,"n_targets":34,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1059,"Elo_lo":1007,"Elo_hi":1109,"n_targets":43,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1026,"Elo_lo":989,"Elo_hi":1062,"n_targets":34,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1025,"Elo_lo":938,"Elo_hi":1097,"n_targets":32,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1024,"Elo_lo":964,"Elo_hi":1093,"n_targets":35,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1019,"Elo_lo":960,"Elo_hi":1078,"n_targets":37,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1006,"Elo_lo":947,"Elo_hi":1067,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":43,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":999,"Elo_lo":964,"Elo_hi":1034,"n_targets":34,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"XT","Elo":997,"Elo_lo":973,"Elo_hi":1021,"n_targets":43,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":995,"Elo_lo":927,"Elo_hi":1059,"n_targets":32,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":980,"Elo_lo":918,"Elo_hi":1046,"n_targets":43,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":980,"Elo_lo":937,"Elo_hi":1023,"n_targets":43,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":976,"Elo_lo":922,"Elo_hi":1035,"n_targets":36,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":973,"Elo_lo":919,"Elo_hi":1027,"n_targets":41,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":966,"Elo_lo":929,"Elo_hi":1005,"n_targets":40,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":958,"Elo_lo":903,"Elo_hi":1011,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":926,"Elo_lo":867,"Elo_hi":981,"n_targets":43,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":869,"Elo_lo":823,"Elo_hi":913,"n_targets":35,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":837,"Elo_lo":791,"Elo_hi":883,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":830,"Elo_lo":775,"Elo_hi":883,"n_targets":43,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n50","cell_label":"p=all, n=50","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":636,"Elo_lo":552,"Elo_hi":710,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1245,"Elo_lo":1176,"Elo_hi":1311,"n_targets":40,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1230,"Elo_lo":1156,"Elo_hi":1295,"n_targets":40,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1218,"Elo_lo":1157,"Elo_hi":1279,"n_targets":29,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1212,"Elo_lo":1130,"Elo_hi":1288,"n_targets":29,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1196,"Elo_lo":1149,"Elo_hi":1240,"n_targets":42,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1191,"Elo_lo":1134,"Elo_hi":1251,"n_targets":40,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1174,"Elo_lo":1126,"Elo_hi":1224,"n_targets":34,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1123,"Elo_lo":1048,"Elo_hi":1200,"n_targets":32,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1122,"Elo_lo":1050,"Elo_hi":1179,"n_targets":13,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1084,"Elo_lo":1035,"Elo_hi":1128,"n_targets":31,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1070,"Elo_lo":1007,"Elo_hi":1137,"n_targets":41,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1055,"Elo_lo":1004,"Elo_hi":1114,"n_targets":35,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1051,"Elo_lo":990,"Elo_hi":1109,"n_targets":42,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1047,"Elo_lo":985,"Elo_hi":1107,"n_targets":42,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1046,"Elo_lo":999,"Elo_hi":1087,"n_targets":37,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1044,"Elo_lo":972,"Elo_hi":1112,"n_targets":42,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1042,"Elo_lo":970,"Elo_hi":1118,"n_targets":30,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1041,"Elo_lo":981,"Elo_hi":1104,"n_targets":36,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1027,"Elo_lo":951,"Elo_hi":1098,"n_targets":31,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1026,"Elo_lo":937,"Elo_hi":1108,"n_targets":31,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1023,"Elo_lo":966,"Elo_hi":1082,"n_targets":42,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1014,"Elo_lo":984,"Elo_hi":1050,"n_targets":36,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1005,"Elo_lo":929,"Elo_hi":1076,"n_targets":41,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1001,"Elo_lo":948,"Elo_hi":1054,"n_targets":39,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1000,"Elo_lo":967,"Elo_hi":1036,"n_targets":39,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":42,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":997,"Elo_lo":946,"Elo_hi":1047,"n_targets":41,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"XT","Elo":984,"Elo_lo":964,"Elo_hi":1004,"n_targets":42,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":951,"Elo_lo":888,"Elo_hi":1011,"n_targets":32,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":905,"Elo_lo":851,"Elo_hi":958,"n_targets":42,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":904,"Elo_lo":856,"Elo_hi":952,"n_targets":34,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":886,"Elo_lo":841,"Elo_hi":927,"n_targets":42,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":860,"Elo_lo":798,"Elo_hi":921,"n_targets":41,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n100","cell_label":"p=all, n=100","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":549,"Elo_lo":457,"Elo_hi":618,"n_targets":42,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1261,"Elo_lo":1195,"Elo_hi":1336,"n_targets":34,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1219,"Elo_lo":1137,"Elo_hi":1302,"n_targets":13,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1209,"Elo_lo":1126,"Elo_hi":1302,"n_targets":26,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1209,"Elo_lo":1152,"Elo_hi":1268,"n_targets":34,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1202,"Elo_lo":1136,"Elo_hi":1267,"n_targets":34,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1193,"Elo_lo":1123,"Elo_hi":1260,"n_targets":34,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1186,"Elo_lo":1092,"Elo_hi":1269,"n_targets":26,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1170,"Elo_lo":1114,"Elo_hi":1229,"n_targets":30,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1160,"Elo_lo":1092,"Elo_hi":1227,"n_targets":22,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1123,"Elo_lo":1055,"Elo_hi":1195,"n_targets":27,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1097,"Elo_lo":1038,"Elo_hi":1156,"n_targets":30,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1081,"Elo_lo":992,"Elo_hi":1175,"n_targets":26,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1077,"Elo_lo":984,"Elo_hi":1160,"n_targets":34,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1069,"Elo_lo":980,"Elo_hi":1156,"n_targets":24,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1064,"Elo_lo":996,"Elo_hi":1133,"n_targets":34,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":1061,"Elo_lo":990,"Elo_hi":1132,"n_targets":34,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1061,"Elo_lo":980,"Elo_hi":1138,"n_targets":32,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1050,"Elo_lo":987,"Elo_hi":1111,"n_targets":34,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"LR","Elo":1046,"Elo_lo":964,"Elo_hi":1124,"n_targets":34,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1046,"Elo_lo":981,"Elo_hi":1112,"n_targets":29,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1041,"Elo_lo":1000,"Elo_hi":1086,"n_targets":28,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1037,"Elo_lo":940,"Elo_hi":1127,"n_targets":24,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1034,"Elo_lo":996,"Elo_hi":1074,"n_targets":28,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1020,"Elo_lo":943,"Elo_hi":1089,"n_targets":34,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1010,"Elo_lo":955,"Elo_hi":1070,"n_targets":28,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1008,"Elo_lo":972,"Elo_hi":1049,"n_targets":30,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":1004,"Elo_lo":928,"Elo_hi":1084,"n_targets":28,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1001,"Elo_lo":936,"Elo_hi":1065,"n_targets":34,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":34,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":993,"Elo_lo":933,"Elo_hi":1053,"n_targets":34,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"XT","Elo":991,"Elo_lo":971,"Elo_hi":1011,"n_targets":34,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":947,"Elo_lo":885,"Elo_hi":1011,"n_targets":34,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":840,"Elo_lo":764,"Elo_hi":910,"n_targets":34,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n200","cell_label":"p=all, n=200","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":485,"Elo_lo":383,"Elo_hi":567,"n_targets":34,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1213,"Elo_lo":1109,"Elo_hi":1333,"n_targets":22,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1192,"Elo_lo":1093,"Elo_hi":1310,"n_targets":14,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1169,"Elo_lo":1096,"Elo_hi":1264,"n_targets":22,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1145,"Elo_lo":1024,"Elo_hi":1243,"n_targets":6,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1142,"Elo_lo":1038,"Elo_hi":1252,"n_targets":22,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1127,"Elo_lo":1033,"Elo_hi":1226,"n_targets":21,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1127,"Elo_lo":1024,"Elo_hi":1238,"n_targets":18,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1118,"Elo_lo":989,"Elo_hi":1246,"n_targets":13,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1103,"Elo_lo":1016,"Elo_hi":1207,"n_targets":19,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1102,"Elo_lo":997,"Elo_hi":1210,"n_targets":22,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1068,"Elo_lo":962,"Elo_hi":1183,"n_targets":14,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1063,"Elo_lo":983,"Elo_hi":1147,"n_targets":22,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1048,"Elo_lo":950,"Elo_hi":1140,"n_targets":21,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1047,"Elo_lo":964,"Elo_hi":1129,"n_targets":22,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1038,"Elo_lo":941,"Elo_hi":1137,"n_targets":21,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1032,"Elo_lo":998,"Elo_hi":1066,"n_targets":21,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1030,"Elo_lo":929,"Elo_hi":1125,"n_targets":22,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1029,"Elo_lo":966,"Elo_hi":1088,"n_targets":21,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1001,"Elo_lo":850,"Elo_hi":1143,"n_targets":22,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":22,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"XT","Elo":996,"Elo_lo":969,"Elo_hi":1020,"n_targets":22,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":987,"Elo_lo":890,"Elo_hi":1084,"n_targets":21,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":982,"Elo_lo":904,"Elo_hi":1058,"n_targets":22,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":981,"Elo_lo":892,"Elo_hi":1071,"n_targets":22,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":979,"Elo_lo":905,"Elo_hi":1052,"n_targets":19,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":975,"Elo_lo":919,"Elo_hi":1037,"n_targets":22,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":963,"Elo_lo":818,"Elo_hi":1108,"n_targets":14,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":952,"Elo_lo":807,"Elo_hi":1130,"n_targets":13,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":931,"Elo_lo":848,"Elo_hi":1010,"n_targets":22,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":921,"Elo_lo":834,"Elo_hi":1018,"n_targets":18,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"LR","Elo":919,"Elo_lo":774,"Elo_hi":1046,"n_targets":22,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":918,"Elo_lo":818,"Elo_hi":1015,"n_targets":22,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":735,"Elo_lo":615,"Elo_hi":837,"n_targets":22,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n500","cell_label":"p=all, n=500","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":363,"Elo_lo":219,"Elo_hi":505,"n_targets":22,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1253,"Elo_lo":1139,"Elo_hi":1385,"n_targets":13,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1211,"Elo_lo":1079,"Elo_hi":1362,"n_targets":17,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1210,"Elo_lo":1127,"Elo_hi":1318,"n_targets":17,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1162,"Elo_lo":1049,"Elo_hi":1285,"n_targets":17,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1114,"Elo_lo":988,"Elo_hi":1241,"n_targets":10,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1114,"Elo_lo":971,"Elo_hi":1278,"n_targets":17,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1114,"Elo_lo":985,"Elo_hi":1253,"n_targets":17,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1085,"Elo_lo":956,"Elo_hi":1233,"n_targets":17,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1085,"Elo_lo":969,"Elo_hi":1225,"n_targets":17,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1060,"Elo_lo":900,"Elo_hi":1233,"n_targets":10,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1053,"Elo_lo":942,"Elo_hi":1169,"n_targets":17,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1015,"Elo_lo":899,"Elo_hi":1133,"n_targets":17,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1010,"Elo_lo":900,"Elo_hi":1117,"n_targets":2,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1001,"Elo_lo":912,"Elo_hi":1091,"n_targets":17,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":17,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT","Elo":997,"Elo_lo":961,"Elo_hi":1036,"n_targets":17,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":993,"Elo_lo":951,"Elo_hi":1032,"n_targets":17,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":984,"Elo_lo":840,"Elo_hi":1136,"n_targets":17,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":983,"Elo_lo":895,"Elo_hi":1060,"n_targets":17,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":979,"Elo_lo":852,"Elo_hi":1116,"n_targets":17,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":978,"Elo_lo":907,"Elo_hi":1048,"n_targets":17,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":978,"Elo_lo":928,"Elo_hi":1026,"n_targets":17,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":968,"Elo_lo":823,"Elo_hi":1115,"n_targets":17,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":957,"Elo_lo":831,"Elo_hi":1089,"n_targets":17,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":945,"Elo_lo":792,"Elo_hi":1136,"n_targets":13,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":927,"Elo_lo":729,"Elo_hi":1129,"n_targets":17,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":923,"Elo_lo":846,"Elo_hi":998,"n_targets":17,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":901,"Elo_lo":698,"Elo_hi":1107,"n_targets":10,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":860,"Elo_lo":739,"Elo_hi":996,"n_targets":17,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":832,"Elo_lo":681,"Elo_hi":965,"n_targets":17,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"LR","Elo":794,"Elo_lo":557,"Elo_hi":988,"n_targets":17,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":719,"Elo_lo":587,"Elo_hi":843,"n_targets":15,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":693,"Elo_lo":572,"Elo_hi":803,"n_targets":17,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full_n1000","cell_label":"p=all, n=1,000","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":320,"Elo_lo":152,"Elo_hi":474,"n_targets":17,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON","Elo":1238,"Elo_lo":1175,"Elo_hi":1309,"n_targets":30,"id":"SELDON","display":"Seldon","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon","description":"Seldon, by Neuralk-AI, is a transformer pretrained only on synthetic tables, with 213.5M classification and 215.5M regression parameters. It compresses cells into row representations and uses labelled training rows as context for classification and regression without updating its weights. This benchmark keeps its released eight-member classification and four-member regression ensembles and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-LARGE","Elo":1222,"Elo_lo":1162,"Elo_hi":1296,"n_targets":41,"id":"KUMO-TABULAR-LARGE","display":"Kumo Tabular (L)","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/nvidia/Kumo-Tabular","description":"Kumo Tabular Large is the released large tabular foundation model from NVIDIA for classification and regression. It uses labelled training rows as context without updating its weights. This benchmark uses its eight-member ensemble, processes members sequentially, and lets the benchmark control feature and sample budgets. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"KUMO-TABULAR-MEDIUM","Elo":1205,"Elo_lo":1143,"Elo_hi":1270,"n_targets":41,"id":"KUMO-TABULAR-MEDIUM","display":"Kumo Tabular (M)","tuned_from":null,"classification_only":false,"source_url":"https://github.com/NVIDIA/structured-data-models","description":"Kumo Tabular, by NVIDIA, learns from synthetic tables before it is applied to a new dataset. It builds representations of columns and rows, then uses labelled examples as context for classification or regression; this benchmark uses the medium-sized model. The released default recipe selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3.5","Elo":1203,"Elo_lo":1158,"Elo_hi":1257,"n_targets":44,"id":"TABPFN-V3.5","display":"TabPFN 3.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3.5, the latest version by Prior Labs, is pretrained on synthetic tables and uses labelled examples as context for new predictions. A single shared checkpoint handles both classification and regression, rather than using separate pretrained models for the two tasks.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":20000,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"SELDON-MINI","Elo":1189,"Elo_lo":1129,"Elo_hi":1256,"n_targets":42,"id":"SELDON-MINI","display":"Seldon-Mini","tuned_from":null,"classification_only":false,"source_url":"https://huggingface.co/Neuralk-AI/seldon-mini","description":"Seldon-Mini is Neuralk-AI's compact tabular foundation model, pretrained only on synthetic tables, with 27.4M classification and 28.4M regression parameters. It uses labelled examples as context without updating its weights. This benchmark keeps its released four-member inference ensemble and lets the benchmark control feature and sample budgets. The released recipe normally selects at most 500 features; this benchmark removes that selection step and evaluates the full feature budget.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"FTFM","Elo":1178,"Elo_lo":1106,"Elo_hi":1259,"n_targets":25,"id":"FTFM","display":"FTFM-1","tuned_from":null,"classification_only":true,"source_url":"https://github.com/machinelearningnuremberg/ftfm-official","description":"FTFM-1 builds cell representations from factorized row and column representations, then predicts classes from labelled training rows supplied as context. This benchmark uses its released eight-member ensemble with sequential views.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"AUTOGLUON","Elo":1173,"Elo_lo":1113,"Elo_hi":1233,"n_targets":43,"id":"AUTOGLUON","display":"AutoGluon","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"AutoGluon automates tabular learning by training multiple models and combining their predictions in ensembles. Here it runs its native extreme preset within a one-hour budget as a separate AutoML reference.","tuning":null,"category":"AutoML","color":"#111827","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"CAUSILO","Elo":1162,"Elo_lo":1099,"Elo_hi":1233,"n_targets":44,"id":"CAUSILO","display":"Causilo","tuned_from":null,"classification_only":false,"source_url":"https://github.com/nums-ai/causilo","description":"Causilo, by Nums AI, is pretrained on synthetic tables and uses labelled examples as context for classification or regression. It processes relationships across rows and columns, combining information about individual features with patterns shared by samples.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABICL","Elo":1161,"Elo_lo":1076,"Elo_hi":1243,"n_targets":4,"id":"TABICL","display":"TabICL v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/soda-inria/tabicl","description":"TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"LIMIX-2","Elo":1157,"Elo_lo":1086,"Elo_hi":1238,"n_targets":21,"id":"LIMIX-2","display":"LimiX2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/limix-ldm-ai/LimiX","description":"LimiX-2 is pretrained on synthetic tables generated from causal models. It learns relationships between features and targets through masked prediction, then uses labelled rows as context to predict classes or numerical values in a new table.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2.5","Elo":1123,"Elo_lo":1061,"Elo_hi":1193,"n_targets":36,"id":"REALTABPFN-V2.5","display":"RealTabPFN 2.5","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 2.5, by Prior Labs, uses a pretrained transformer to make predictions from labelled examples, extending the v2 family to wider tables. Its default classifier includes pretraining on real datasets, while its default regressor is pretrained on synthetic data only.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABFM","Elo":1119,"Elo_lo":1035,"Elo_hi":1215,"n_targets":14,"id":"TABFM","display":"TabFM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/google-research/tabfm","description":"TabFM is Google's pretrained model for classification and regression on tables containing numerical and categorical columns. It predicts from labelled training rows supplied as context, using sampled groups of rows and features when needed, rather than updating its model weights for each new dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2000,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-V3","Elo":1099,"Elo_lo":1038,"Elo_hi":1168,"n_targets":33,"id":"TABPFN-V3","display":"TabPFN 3","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN 3, by Prior Labs, is a transformer pretrained on synthetic prediction tasks. For classification or regression on a new table, it uses the labelled training rows as examples to guide its predictions rather than learning the task from scratch.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":10000,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABM","Elo":1097,"Elo_lo":1038,"Elo_hi":1159,"n_targets":42,"id":"TABM","display":"TabM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/yandex-research/tabm","description":"TabM, by Yandex Research, trains a group of multilayer perceptrons together for tabular classification or regression, with most of their weights shared. Each input produces several predictions that are combined, giving the model an ensemble structure without storing and training a full set of separate networks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"CAT","Elo":1087,"Elo_lo":1042,"Elo_hi":1136,"n_targets":39,"id":"CAT","display":"CatBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/catboost/catboost","description":"CatBoost builds an ensemble of decision trees through gradient boosting, adding trees that improve the current predictions. It supports numerical and categorical features for classification and regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE-5K-NE3","Elo":1076,"Elo_lo":1003,"Elo_hi":1147,"n_targets":30,"id":"TABPFN-WIDE-5K-NE3","display":"TabPFN Wide 5k (ne3)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This TabPFN-Wide configuration uses the wide-v2-5k checkpoint, pretrained for classification with many features and few samples. It averages three inference variants while keeping the checkpoint's original feature grouping.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABDPT","Elo":1073,"Elo_lo":1022,"Elo_hi":1132,"n_targets":44,"id":"TABDPT","display":"TabDPT","tuned_from":null,"classification_only":false,"source_url":"https://github.com/layer6ai-labs/TabDPT-inference","description":"TabDPT is a tabular foundation model by Layer6 AI pretrained on real datasets using self-supervised prediction tasks. For a new classification or regression problem, it retrieves relevant labelled rows and uses them as context to predict new rows without updating its weights on that dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":2500,"training_data_overlap":true},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"LR-TUNED","Elo":1060,"Elo_lo":998,"Elo_hi":1124,"n_targets":41,"id":"LR-TUNED","display":"Logistic Regression (tuned)","tuned_from":"LR","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This is the same linear model as LR: logistic regression for classification and ridge regression for continuous targets. A small predefined grid selects regularization strength on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"C":[0.01,0.1,10,100,0.001,0.0001]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH","Elo":1053,"Elo_lo":993,"Elo_hi":1113,"n_targets":44,"id":"NN_TORCH","display":"MLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This multilayer perceptron uses AutoGluon's PyTorch implementation for tabular classification and regression. It passes the input features through fully connected hidden layers with nonlinear activations to learn the prediction function.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"GBM","Elo":1047,"Elo_lo":998,"Elo_hi":1103,"n_targets":44,"id":"GBM","display":"LightGBM","tuned_from":null,"classification_only":false,"source_url":"https://github.com/lightgbm-org/LightGBM","description":"LightGBM is a gradient boosting method that builds decision trees using binned feature values. It grows each tree by splitting the leaf with the largest reduction in the training objective.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"NN_TORCH-TUNED","Elo":1044,"Elo_lo":996,"Elo_hi":1091,"n_targets":40,"id":"NN_TORCH-TUNED","display":"MLP (tuned)","tuned_from":"NN_TORCH","classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"This variant uses AutoGluon's PyTorch multilayer perceptron, selecting its depth, hidden layer width, learning rate and dropout from a small grid. Selection uses a holdout drawn only from the training data, followed by a refit of the selected configuration.","tuning":{"grid":[{"dropout_prob":[0.0,0.2],"hidden_size":[64,128],"learning_rate":[0.0003,0.001],"num_layers":[2,4]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"XT-TUNED","Elo":1040,"Elo_lo":1000,"Elo_hi":1080,"n_targets":38,"id":"XT-TUNED","display":"Extra Trees (tuned)","tuned_from":"XT","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Extra Trees model as XT for classification or regression. A small predefined grid varies candidate features per split, minimum leaf size and bootstrap sampling on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1],"min_samples_leaf":[1,5]},{"bootstrap":[false]},{"bootstrap":[false],"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"REALMLP","Elo":1031,"Elo_lo":966,"Elo_hi":1092,"n_targets":43,"id":"REALMLP","display":"RealMLP","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dholzmueller/pytabkit","description":"RealMLP is a multilayer perceptron trained on each dataset for classification or regression. It combines numerical preprocessing with a carefully chosen network and training setup, using defaults selected across a broad collection of tabular tasks.","tuning":null,"category":"Deep Learning","color":"#dc2626","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"REALTABPFN-V2","Elo":1024,"Elo_lo":955,"Elo_hi":1092,"n_targets":40,"id":"REALTABPFN-V2","display":"RealTabPFN v2","tuned_from":null,"classification_only":false,"source_url":"https://github.com/PriorLabs/TabPFN","description":"TabPFN v2, by Prior Labs, is a pretrained transformer for classification and regression on tables. It takes labelled training rows as context and predicts new rows without training a separate neural network from scratch for each dataset.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"RF-TUNED","Elo":1019,"Elo_lo":992,"Elo_hi":1050,"n_targets":39,"id":"RF-TUNED","display":"Random Forest (tuned)","tuned_from":"RF","classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"This uses the same Random Forest as RF, combining trees for classification or regression. A small predefined grid selects the number of candidate features per split and minimum leaf size on a training-only holdout, then the selected model is refitted on the full training fold.","tuning":{"grid":[{"max_features":["log2",0.1,"sqrt",0.5],"min_samples_leaf":[1,5]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"XGB-TUNED","Elo":1018,"Elo_lo":976,"Elo_hi":1061,"n_targets":34,"id":"XGB-TUNED","display":"XGBoost (tuned)","tuned_from":"XGB","classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"This variant uses the same XGBoost tree boosting model, with settings selected from a small grid on a holdout drawn only from the training data. The grid varies tree depth, learning rate, feature sampling and regularization before refitting the selected configuration.","tuning":{"grid":[{"learning_rate":[0.03,0.1],"max_depth":[3,6],"n_estimators":[500]},{"colsample_bytree":[0.1,0.5],"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500]},{"learning_rate":[0.03,0.1],"max_depth":[1,2],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"min_child_weight":[5,20],"n_estimators":[500]},{"learning_rate":[0.1],"max_depth":[3,6],"n_estimators":[500],"reg_lambda":[10,100]}],"protocol":"holdout-grid-v1","search_fraction":0.8,"seed":42,"selection":"inner_holdout","validation_fraction":0.2,"budget_seconds":3600},"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"TABPFN-WIDE","Elo":1015,"Elo_lo":928,"Elo_hi":1098,"n_targets":30,"id":"TABPFN-WIDE","display":"TabPFN Wide (8k)","tuned_from":null,"classification_only":true,"source_url":"REMOVED","description":"TabPFN-Wide continues pre-training TabPFN 2 on synthetic tables with many features and few samples, a setting common in omics. This classification model uses the wide-v2-8k checkpoint to work with wider inputs than the original model.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"RF","Elo":1000,"Elo_lo":1000,"Elo_hi":1000,"n_targets":44,"id":"RF","display":"Random Forest","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Random Forest grows decision trees on bootstrap samples of the training data, considering a random subset of features at each split. It averages their class probabilities for classification or their numerical predictions for regression.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"LR","Elo":997,"Elo_lo":919,"Elo_hi":1072,"n_targets":44,"id":"LR","display":"Logistic Regression","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Logistic regression learns a weighted combination of the input features and turns it into class probabilities. For continuous targets, this entry uses ridge regression, a linear model with a penalty that keeps its coefficients from becoming too large.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"XGB","Elo":988,"Elo_lo":942,"Elo_hi":1037,"n_targets":44,"id":"XGB","display":"XGBoost","tuned_from":null,"classification_only":false,"source_url":"https://github.com/dmlc/xgboost","description":"XGBoost combines decision trees through gradient boosting, adding trees to improve the ensemble's predictions. Its training objective includes penalties on tree complexity, alongside the loss for classification or regression.","tuning":null,"category":"Gradient Boosting","color":"#d97706","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"XT","Elo":984,"Elo_lo":959,"Elo_hi":1007,"n_targets":44,"id":"XT","display":"Extra Trees","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"Extra Trees averages the predictions of many decision trees. Each tree considers randomly chosen features and split thresholds, adding more randomness than a conventional random forest when learning a classification or regression task.","tuning":null,"category":"Tree-based","color":"#059669","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"GENEICL","Elo":972,"Elo_lo":895,"Elo_hi":1043,"n_targets":42,"id":"GENEICL","display":"GeneICL","tuned_from":null,"classification_only":false,"source_url":"https://github.com/BoevaLab/GeneICL","description":"GeneICL is built for prediction from bulk gene expression profiles. It compresses the gene measurements with principal component analysis, then uses a pretrained recurrent network to learn from labelled examples without updating its weights for each new task.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"KNN","Elo":856,"Elo_lo":799,"Elo_hi":909,"n_targets":43,"id":"KNN","display":"KNN","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"KNN predicts from the training examples closest to a new sample in feature space. The benchmark uses equally weighted neighbors, taking their class votes for classification or averaging their target values for regression.","tuning":null,"category":"Traditional ML","color":"#2563eb","regular_max_features":null,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"MITRA","Elo":814,"Elo_lo":718,"Elo_hi":898,"n_targets":28,"id":"MITRA","display":"MITRA","tuned_from":null,"classification_only":false,"source_url":"https://github.com/autogluon/autogluon","description":"MITRA is a pretrained model for tabular classification and regression that learns from a mixture of synthetic datasets generated by different rules. It uses labelled examples as context for new predictions, and AutoGluon's implementation also adapts its pretrained weights to the target dataset through fine-tuning.","tuning":null,"category":"Tabular Foundation","color":"#7c3aed","regular_max_features":500,"training_data_overlap":false},{"cell":"cap_full","cell_label":"p=all, n=all","domain":"all","metric":"f1_macro","model_id":"DUMMY","Elo":361,"Elo_lo":254,"Elo_hi":446,"n_targets":44,"id":"DUMMY","display":"Constant","tuned_from":null,"classification_only":false,"source_url":"https://github.com/scikit-learn/scikit-learn","description":"The constant baseline ignores the input features. It returns the training set's class proportions for classification or its mean target value for regression, showing what can be predicted without using the measurements.","tuning":null,"category":"Baseline","color":"#9ca3af","regular_max_features":null,"training_data_overlap":false}]}}}
