Apply regression target transforms to the unnormalized target [RES-2639]

[bejaeger]

Closes RES-2639.

Problem

The regressor z-normalises the target before the ensemble preprocessing, so every entry of REGRESSION_Y_PREPROCESS_TRANSFORMS acted on standardized values rather than on the target. 1_plus_log was log1p of a z-score, not of the target — undefined wherever the z-score drops below -1, which is 15% of the rows for a symmetric target, and those NaNs reached the model unguarded because the log presets have no trailing scaler.

Change

wrap_target_transform composes each transform into a pipeline that

1. undoes the z-normalisation, so the transform sees the target in its original units,
2. applies the transform,
3. standardizes the result — the scale the bar distribution's borders assume.

The pipeline's input and output stay in the z-normalised space, so nothing downstream moves: inverse_transform maps the model's bar-distribution borders straight back into z-space, exactly as an unwrapped transform did, and the ±1e3 sanity limits applied to those borders keep the z-units they were tuned in. That is why predict, predict_batched and the border mapping are untouched.

The fine-tuning data pipeline re-splits and re-normalises per split, so it rebinds the statistics before fitting. "none" is now treated as no transform at all, since composing the identity would only add float noise.

Also fixes the docstring of _transform_labels_one, which called its input "unprocessed" (the second item on the ticket).

Impact

Changes the v2 / v2.5 / v3 defaults, whose ensembles use safepower for half of their estimators — up to ~4% relative prediction change on california housing. Bit-identical for the v2.6 default (("none",)), for n_estimators=1 on any version, and for every estimator whose target transform is None.

Tests

New tests/test_preprocessing/test_target_transform.py (17 cases) plus 4 regressor-level tests; two of them fail on main and pass here. Full local sweep of regressor + classifier interface, preprocessing, consistency, finetuning-regressor, save/load and inference-tuning suites passes.

🤖 Generated with Claude Code

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want fixes drafted automatically? Bugbot Autofix can create code changes for findings. A team admin can enable Autofix in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit bf24f1d. Configure here.}

[cursor]

Finetuning skips fitted target transforms

Medium Severity

RegressorBatch still ships list(conf) for border mapping, not the configs returned on ensemble_members after fit_transform_ensemble_members. With n_preprocessing_jobs > 1, wrap_target_transform's outer StandardScaler is fitted only on the worker copy, so finetuning inverse_transform on bar borders can hit an unfitted pipeline or decode with the wrong scale. predict_batched already takes member configs for this reason.

Additional Locations (1)

• src/tabpfn/finetuning/data_util.py#L437-L442

 

^{Triggered by learned rule: Use executor_.ensemble_members configs, not ensemble_configs_, for fitted transforms}

^{Reviewed by Cursor Bugbot for commit bf24f1d. Configure here.}

[bejaeger]

Superseded by the #1198 → #1199 stack, which removes the estimator's up-front z-normalisation of the target instead of composing it into each target pipeline. Same semantics as this PR (verified to 1e-7 relative on five datasets, so the TabArena validation transfers), without the baked-in statistics or the rebinding they forced on the fine-tuning path.