Drive splitters with stock scikit-learn CVs via groups=

[Copilot]

Summary

One mechanism, every evaluation, stock scikit-learn CVs only. Pass any standard scikit-learn cross-validator as cv_class and steer the folds from the metadata via groups= and/or callable cv_kwargs — no custom splitter classes, no metadata-aware detection, no subclassing.

CrossSubjectEvaluation(...)                                          # leave-one-subject-out (default, unchanged)
CrossSubjectEvaluation(..., cv_class=LeaveOneGroupOut, groups=["subject", "session"])
CrossSubjectEvaluation(..., cv_class=PredefinedSplit,
    cv_kwargs={"test_fold": lambda md: np.where((md.subject == 2) & (md.session == "1"), 0, -1)})  # single-target
WithinSubjectEvaluation(..., cv_class=GroupKFold, groups="run")      # the "special detail" case

What changed

• groups= — a metadata column name, a list of column names (compound key, e.g. ["subject", "session"]), or a callable metadata -> array. Available on CrossSubject, CrossDataset, CrossSession (the inner held-out axis), and the Within* splitters (optional, default None = legacy behaviour). Resolved from the metadata and handed straight to the stock cv_class.
• Callable cv_kwargs — any cv_kwargs value that is a callable is resolved against the metadata at split time, e.g. PredefinedSplit.test_fold for a single-target fold.
• CrossSubjectSplitter and CrossDatasetSplitter are kept as separate classes (leave-one-subject-out vs leave-one-dataset-out are semantically distinct). CrossDatasetSplitter's group_column still works but is deprecated in favour of groups.

Defaults reproduce the current behaviour exactly (backward compatible).

[chatgpt-codex-connector]

💡 Codex Review

moabb/test_save_path/fitted_model_0.pkl

Line 1 in 5b80814

���w� sklearn.dummy���DummyClassifier���)��}�(��strategy���prior��random_state�N��constant�N��_sklearn_version���1.9.0�ub.

Remove the committed generated model artifact

Because this pickle is already present in checkouts, TestUtilEvaluation.test_save_model_cv can pass even when save_model_cv() fails to create anything—the assertion only checks that test_save_path/fitted_model_0.pkl exists after the call. This masks regressions in model saving and leaves a generated test artifact in the source tree; remove it and have the test create/clean the path.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".