Add TabFM and fix foundation-model integration bugs

[andreasgoethals]

Add TabFM and fix foundation-model integration bugs

Summary

• Add tabfm, an optional Google Research TabFM v1.0.0 method wrapper using the upstream tabfm[pytorch] package.
• Register tabfm in the unified method registry and add matching default / opt-space configs.
• Document TabFM in the README, including the optional install command and the upstream non-commercial model-weight license caveat.
• Fix TALENT/configs/opt_space/tabptm.json, whose top-level key incorrectly pointed to tabpfn.
• Fix shared preprocessing edge cases:
  • unknown categorical values now map to the training-column mode as documented;
  • constant regression targets no longer divide by zero during label normalization.
• Fix Python API argument handling so TALENT.run(..., args=...) respects supplied seed, seed_num, tune, n_trials, and tune_metric fields unless explicitly overridden.
• Make TabPTM temporary tensors follow the active device instead of hardcoding CUDA.

Notes

• tabfm is optional and not added as a hard dependency. Users should install it with:

  pip install -U "tabfm[pytorch]"

• TabFM v1.0.0 classification is limited to 10 classes by the upstream model. The wrapper raises a clear ValueError before loading the model if a dataset exceeds that limit.

• The wrapper uses the PyTorch backend because TALENT already centers on PyTorch and the upstream TabFM package exposes tabfm_v1_0_0_pytorch.

• The upstream TabFM source code is Apache 2.0, while the published model weights are under the TabFM Non-Commercial License.

Validation

• python -m compileall -q TALENT test
• git diff --check
• Registry/config consistency check:
  • 56 registered methods;
  • no missing config files;
  • no mismatched config top-level keys.
• Focused preprocessing checks for categorical unknown replacement and constant regression target normalization.
• API behavior check confirming args.tune=True is now honored by TALENT.run(..., args=...).

Follow-Up

• Full runtime smoke tests for tabfm require installing tabfm[pytorch] and downloading the upstream Hugging Face weights.
• Consider adding a lightweight CI test that checks registry/config consistency for every registered method.

[Copilot]

Pull request overview

This PR adds a new optional TabFM method integration to TALENT’s unified method registry/config system, while also fixing several integration and preprocessing edge cases that affect foundation-model wrappers and the Python API.

Changes:

• Added a tabfm method wrapper (optional dependency) and registered it with matching default + opt-space configs.
• Fixed preprocessing edge cases (unknown categorical replacement and constant regression-target normalization).
• Fixed Python API TALENT.run(..., args=...) override/default handling and improved device handling in TabPTM temporary tensors.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
TALENT/model/methods/tabptm.py	Makes temporary tensors device-aware (no hardcoded .cuda()), plus trailing whitespace cleanup.
TALENT/model/methods/tabfm.py	Adds a new TabFM wrapper (optional tabfm[pytorch]) for zero-shot classification/regression.
TALENT/model/method_registry.py	Registers tabfm in the unified method registry with constraints and notes.
TALENT/model/lib/data.py	Fixes unknown-categorical replacement logic and guards against divide-by-zero in regression label normalization.
TALENT/configs/opt_space/tabptm.json	Fixes incorrect top-level key (tabpfn → tabptm).
TALENT/configs/opt_space/tabfm.json	Adds TabFM opt-space config stub + general defaults.
TALENT/configs/default/tabfm.json	Adds TabFM default config stub + general defaults.
TALENT/api.py	Updates run() kwarg/args precedence logic for seeds and tuning fields.
readme.md	Documents TabFM as an optional method and updates the foundation-model cap note + acknowledgements.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.