Tabular foundation models

Which benchmark datasets could each model have seen during training? Every coloured square links to a verbatim quote from the model's paper, model card or repository. Grey means not enough has been disclosed to tell.

Contamination map

Rows are models, grouped by family; columns are benchmark datasets. Hover a square for the reason, click it for the evidence. Empty outlines mark benchmarks a model cannot take (e.g. image benchmarks for text-only models).

Benchmarks

Models

ModelOrganisationReleasedDisclosureDetermined
LimiX Stable AI Technology Co., Ltd. (LimiX team) 2025-08-29 Full list 100%
LimiX-2 Stable AI Technology Co., Ltd. (LimiX team) 2026-09-16 Full list 100%
Mitra Amazon 2025-06-22 Full list 100%
Mitra-v2 Amazon 2026-09-03 Full list 100%
TabDPT Layer 6 AI (TD Bank) 2024-10-23 Full list 13%
TabDPT v1.3 Layer 6 AI (TD Bank) 2026-09-08 Vague 0%
TabDPT-Turbo (TabDPT v1.2) Layer 6 AI (TD Bank) 2026-06-17 Vague 100%
TabICL Inria (SODA team) 2025-02-08 Full list 100%
TabICL v1.1 Inria (SODA team) — Full list 100%
TabICLv2 Inria (SODA team) 2026-02-11 Full list 100%
Real-TabPFN Prior Labs 2025-07-05 Full list 5%
Real-TabPFN-2.5 Prior Labs 2025-11-11 Full list 100%
TabPFN v2 Prior Labs / University of Freiburg 2025-01-08 Full list 100%
TabPFN-2.5 Prior Labs 2025-11-11 Full list 100%
TabPFN-2.6 Prior Labs 2026-03-24 Full list 100%
TabPFN-3 Prior Labs 2026-05-13 Full list 100%
TabPFN-3.5 Prior Labs 2026-09-15 Full list 100%

Datasets