Is your feature request related to a problem? Please describe.
LazyPredict benchmarks classical ML models, but there's no way to compare them against pretrained tabular foundation models, which often perform strongly on small/medium datasets - users have to leave LazyPredict to test them.
Describe the solution you'd like
Integrate Google's TabFM as an optional model, following the existing XGBoost/PerpetualBooster pattern: a thin wrapper that lazily loads the pretrained checkpoint at fit() time (TabFM's classes aren't zero-arg constructible), a guarded import so nothing changes without tabfm installed, plus skip-if-missing tests and docs. TimesFM on the forecasting side is direct precedent.
Describe alternatives you've considered
TabPFN (on PyPI, could be a follow-up) and pytorch-tabnet (not pretrained, adds little over existing boosters).
Additional context
TabFM is sklearn-compatible but Python ≥ 3.11 and not on PyPI (from-source install, so no pip extra). I have a working implementation (wrapper + registration + tests + docs) and would be happy to open a PR.
Is your feature request related to a problem? Please describe.
LazyPredict benchmarks classical ML models, but there's no way to compare them against pretrained tabular foundation models, which often perform strongly on small/medium datasets - users have to leave LazyPredict to test them.
Describe the solution you'd like
Integrate Google's TabFM as an optional model, following the existing XGBoost/PerpetualBooster pattern: a thin wrapper that lazily loads the pretrained checkpoint at fit() time (TabFM's classes aren't zero-arg constructible), a guarded import so nothing changes without tabfm installed, plus skip-if-missing tests and docs. TimesFM on the forecasting side is direct precedent.
Describe alternatives you've considered
TabPFN (on PyPI, could be a follow-up) and pytorch-tabnet (not pretrained, adds little over existing boosters).
Additional context
TabFM is sklearn-compatible but Python ≥ 3.11 and not on PyPI (from-source install, so no pip extra). I have a working implementation (wrapper + registration + tests + docs) and would be happy to open a PR.