New datasets
Added the Tox21 SR-MMP and NR-AhR molecular property tasks and the ST001000 stool metabolomics dataset comparing Crohn's disease with ulcerative colitis.
Changelog
Benchmark updates, versioned over time.
Material changes to the benchmark, public artifacts, and website are recorded here.
Failure Map and MLP
Added a failure map view that visualizes the failure rates at different feature and sample budgets. A tuned version of MLP was also added.
LimiX-2 and GeneICL
Added LimiX-2 and GeneICL models.
New (and tuned) models
Added Causalio, Kumo, TabPFN 3.5 and tuned variants of classical ml models.
Probability metrics
Probability metrics retain the complete labelled class space. ROC-AUC is undefined when a test fold lacks a task class; those folds are omitted from ROC-AUC Elo. Rebuilding reports can therefore change ROC-AUC rankings and published numbers. Log loss uses every probability column without dropping absent test classes.
Model cards
Added detailed visualisations of each model's strengths and weaknesses across dataset modalities and training budgets.
Dataset explorer
Added improved per-dataset model comparisons and learning curves across feature and sample budgets.
Initial release
Published the first TabBench-Bio benchmark website, interactive results, dataset registry, and versioned result artifacts.