Skip to content

Remove the all-missing column workaround now fixed in TabPFN - #66

Draft
adrian-prior wants to merge 2 commits into
mainfrom
adrian/drop-all-missing-column-workaround
Draft

adrian-prior wants to merge 2 commits into
mainfrom
adrian/drop-all-missing-column-workaround

Conversation

@adrian-prior

@adrian-prior adrian-prior commented Oct 1, 2026 •

Copy link
Copy Markdown
Collaborator

Generated by Claude Code

#54 dropped training columns that are entirely missing, so that a string in such a column at predict could not crash TabPFN (#51). TabPFN handles this itself from 9.1.0 (PriorLabs/TabPFN#1329), so this PR removes the workaround from fit_tfm and replaces its tests with one that checks an all-missing column reaches the estimator at fit and predict. The rdblearn extra now requires tabpfn>=9.1, and uv.lock moves from 9.0.0 to 9.2.0.

Not covered by this PR: tabpfn-rel also fits through fit_tfm. Its local extra allows tabpfn 9.0.0, which still crashes without the workaround, and its api extra was not tested.

Verification

CI does not install the rdblearn extra, so the checks below with real TabPFN ran by hand on CPU.

  • fit_tfm and predict_tfm with the rdblearn recipes (TabPFN v2 and v2.5), binary classification and regression. The training column was entirely missing as float NaN, entirely missing as object NaN, or populated in the frame but entirely missing in the sampled context rows. The column held strings (including "") at predict.
    • tabpfn==9.1.0 and tabpfn==9.2.0: all 12 cases pass. Predictions are identical to a fit without the column (maximum absolute difference 0), with ROC AUC 0.99 and R² 0.97 or higher on synthetic data.
    • tabpfn==9.0.0 (control, installed without the extra): all 12 cases fail with ValueError: could not convert string to float.
  • The same cases with the local v3 recipe of tabpfn-rel: pass on 9.1.0 and 9.2.0, fail on 9.0.0.
  • The rdblearn code paths that touch the TabPFN API (create_default_for_version, the prediction batch setting, median regression output) run on 9.1.0 and 9.2.0. A full rdblearn benchmark run was not repeated.
  • The CI test job, run locally on Python 3.11 and 3.12: ruff format --check, ruff check, pytest (526 passed, 18 skipped, before the new test; test_tfm.py with it: 9 passed), uv build --all-packages, workflows/verify_distributions.py. pre-commit run --all-files passes.
Details
  • fit_tfm matches its state before Drop all-missing training columns before fitting tabular models #54.
  • The new test also checks that predict_tfm passes columns in training order; it fails when that reindex is removed.
  • The README and docs/licensing.md no longer say rdblearn allows 8.x. They now name tabpfn-rel-api as the extra that still allows it, through tabpfn-extensions>=0.6.3, which requires tabpfn>=8.4.0.
  • uv.lock was regenerated with uv 0.12.8; the only changes are the tabpfn specifier and version.
  • The lock moves every extra from tabpfn 9.0.0 to 9.2.0. 9.1.0 also ships other fixes that change predictions, so reruns of results recorded on 9.0.0 (baseline_results/README.md) can differ.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PivQFscAkcPNjCnGHYkCJF

adrian-prior and others added 2 commits October 1, 2026 14:12
TabPFN restores the fit-time value of columns that are constant at fit
(PriorLabs/TabPFN#1329), so fit_tfm no longer needs to drop them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
TabPFN 9.1 handles a column that is entirely missing at fit and holds
strings at predict, so fit_tfm no longer needs to drop such columns.
Relock to tabpfn 9.2.0, correct the tabpfn license notes, and test that
an all-missing column reaches the estimator in training column order.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PivQFscAkcPNjCnGHYkCJF
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant