Tabular / Benchmark

TabArena TabArena-v0.1

Living benchmark for predictive machine learning on small-to-medium IID tabular data: 51 manually curated classification and regression datasets (500-250,000 training samples), hosted as OpenML tasks (dataset ids 46904-46980, task ids 363612-363712).

HomepageLeaderboard

Test split: Outer evaluation splits are fixed OpenML tasks: 3-fold outer cross-validation (stratified by class for classification), repeated 10 times (30 splits) for datasets with fewer than 2,500 samples and 3 times (9 splits) otherwise. Models are tuned/validated with 8-fold inner cross-validation on each outer training fold. TabArena-Lite uses only the first outer split of every dataset. Every row of a dataset appears in some outer test fold, so any pre-training on the dataset overlaps test data.