build(deps): bump tabpfn from 8.5.0 to 9.0.0

[dependabot]

Bumps tabpfn from 8.5.0 to 9.0.0.

Release notes

Sourced from tabpfn's releases.

v9.0.0

Breaking Changes

• eval_metric="roc_auc" combined with tune_decision_thresholds=True now raises instead of silently optimizing balanced accuracy. The threshold search scores thresholded labels, and the ROC AUC of a single operating point equals (TPR + TNR) / 2, so the two metrics produced identical thresholds. Pass eval_metric="balanced_accuracy" to keep the previous behaviour. (#1153)
• eval_metric="log_loss" combined with tune_decision_thresholds=True now raises instead of silently optimizing accuracy. The threshold search scores hard labels, and log loss over clipped hard labels is a constant multiple of the error rate. Use calibrate_temperature=True, which optimizes log loss directly. (#1239)
• TabPFN-3.5 is the default model: TabPFNClassifier() and TabPFNRegressor() now load ModelVersion.V3_5 unless a model_path or another version is given. The previous default, TabPFN-3, stays available through create_default_for_version(ModelVersion.V3). (#1273)

Added

• InferenceConfig.SOFTMAX_TEMPERATURE lets a checkpoint carry the softmax temperature it was trained for. softmax_temperature now defaults to "auto", meaning "use the checkpoint's value"; passing a float still overrides it for every model, though naming a temperature through both softmax_temperature and inference_config now raises. Existing checkpoints lack the field and fall back to its default of 0.9, the value the interface applied before, so nothing changes for them. Ensembling checkpoints that declare different temperatures now raises and asks for one to be passed explicitly. (#1220)
• InferenceConfig.N_ESTIMATORS lets a checkpoint declare how many estimators to run, and means exactly what the n_estimators argument means: an int is used as given and never raised for feature coverage, "auto" (the default) means 8 and may be raised. n_estimators="auto" now takes the checkpoint's value, so checkpoints without the field behave as before. An explicit n_estimators still wins; naming a count both ways, or ensembling checkpoints that disagree on it, raises. (#1230)
• inference_config={"TRANSFORM_DATES": True} expands a column holding a genuine datetime dtype (datetime64, tz-aware, or period) into calendar features: the year, the day of year, the seconds since the epoch, and the cyclical month, day and weekday pairs, plus the time of day when the column carries one. Off by default, in which case such a column is refused with a TabPFNValidationError naming it. Not run by the fine-tuning estimators. (#1236)
• inference_config={"TRANSFORM_TEXT": True} expands a text column into numeric features via skrub.StringEncoder: tf-idf over its character n-grams, reduced by a truncated SVD to at most TEXT_N_COMPONENTS features (30 by default). A column is text when it has pandas' string dtype, in any storage, or a pyarrow string dtype, more than MIN_CARDINALITY_FOR_TEXT distinct values, and is not declared categorical; an object or category column is never expanded. An unseen string is encoded by the character n-grams it shares with the training column; a missing value, or a string sharing none, becomes an all-zero row. Off by default, in which case nothing changes: such a column is ordinal-encoded as a high-cardinality category and fit() warns about it by name, as before, the warning now pointing at TRANSFORM_TEXT. Fitted estimators expose categorical_features_indices_, the declared categorical positions in the validated input after any date or text expansion. Not run by the fine-tuning estimators. (#1240)
• Add InferenceConfig.SAMPLE_SUBSAMPLING_METHOD to choose how rows are drawn per estimator when SUBSAMPLE_SAMPLES is set: "balanced", "stratified", or the new "majority_downsample", which groups rows by exact target value, keeps every row outside the single most frequent group, and fills the remaining budget from that majority group. This keeps the whole minority class for imbalanced classification and downsamples the repeated value for zero-inflated regression targets. Targets without a unique majority warn and fall back to the task's default sampling method. The default "auto" keeps the previous behavior (stratified for classification, balanced for regression). TabPFNClassifier.fit_with_differentiable_input now also uses stratified row subsampling, matching fit. (#1253)

Changed

• clean_data no longer casts a numeric input to float64; the preprocessing pipeline widens its own copy instead, cutting peak host RAM on a 1M x 2000 float32 fit + predict from 26.1 GB to 15.4 GB with identical outputs. (#1218)
• matplotlib is now a direct dependency instead of behind the optional viz extra (which has been removed), since skrub (a new direct dependency, used by upcoming work) already requires it transitively. (#1222)
• Widen the skrub dependency from the 0.10 series to 0.7 or newer. (#1231)
• The out-of-memory error now reports the feature count TabPFN actually ran on rather than the width of the frame passed to fit, which can be narrower when date expansion is on. (#1236)
• Dates are recognized only from a genuine datetime dtype, never by parsing string columns with pd.to_datetime(): a string column that merely looks like a date is read as a plain category or text. The never-consumed FeatureModality.DATE has been removed. (#1236)
• load_fitted_tabpfn_model and the load_from_fit_state class methods now default to device="auto" instead of device="cpu", so a GPU-fitted estimator no longer silently reloads onto CPU. Pass device="cpu" for the old behavior. (#1243)
• Declared categoricals are taken at face value: a string column listed in categorical_features_indices is categorical at any cardinality, never text, and a pandas category column now counts as declared (its position is merged into categorical_features_indices_). MIN_CARDINALITY_FOR_TEXT only applies to undeclared string columns. (#1245)
• Clarify estimator usage guidance for TabPFN-3 and later versions, covering dataset capacity, subsampling, and automatic handling of categorical and missing feature values without manual preprocessing. (#1247)
• Relicense the TabPFN code under standard Apache 2.0, removing the additional attribution requirement from the code license. Model weight licenses are unchanged: TabPFN-2 retains its existing Prior Labs License, and TabPFN-2.5, TabPFN-2.6, and TabPFN-3 remain non-commercial. (#1271)

Fixed

• • Fixed importance-based feature subsampling drawing every estimator's non-top features independently instead of from a shared round-robin pool, which could leave features unseen by any ensemble member even when the combined feature budget covered them all. Results change for fits using gini_feature_importance (directly or via "auto" on large, wide datasets). (#1141)
• Fix decision-threshold tuning assigning an arbitrary threshold to classes with no rows in the tuning holdout, which heavily boosted the class for eval_metric='roc_auc' and suppressed it for the other metrics. Such classes now receive a neutral threshold. (#1152)
• Cut TorchSquashingScaler's peak device memory from 13.0x its input to 0.4x for fit and 5.8x to 1.0x for transform, with byte-identical outputs, so large fits no longer exhaust VRAM. (#1223)
• Above 100k rows, importance-based feature subsampling gave every estimator its own feature ranking and its own sampling pool, so some features ended up in no ensemble member at all even when the combined feature budget covered them. A single shared ranking now restores full coverage, and replaces up to n_estimators LightGBM fits with one. Results change for gini_feature_importance fits (directly or via "auto") above 100k rows. (#1229)
• Prevent batched predictions from depending on other datasets' preprocessed feature widths. (#1233)
• get_embeddings now cleans categorical columns the way predict does, from the fitted feature schema rather than from the indices declared in categorical_features_indices. (#1236)
• TabPFNRegressor now reports the dataset sizes in its out-of-memory error. Only TabPFNClassifier passed them on, so the regressor's message omitted the "Your sizes" line entirely. (#1236)
• Fitting or predicting on a DataFrame with a genuine datetime column beside any other dtype, or with a timedelta64 or period column, no longer aborts with numpy's dtype-promotion error or Unknown dtype: period[D]. A datetime or period column is refused with a clear error, or expanded with TRANSFORM_DATES; a duration becomes its length in seconds. (#1236)
• Fixed TorchQuantileTransformer's sklearn-equivalence test failing on torch 2.14. torch 2.14 derives the sample index for a quantile in double precision, so the transformer's float32 quantile positions can land just short of an exact order statistic and get interpolated instead, shifting transformed values by up to 1.1e-4 where two neighbouring samples are close together. The test's tolerance now accommodates that; the transform still runs in float32, since computing the fit in float64 would cost performance in the pipeline for a difference this small. (#1238)
• Fixed saving a fitted estimator whose device is a torch.device (or a sequence of them), which raised TypeError: Object of type device is not JSON serializable. (#1243)
• numpy arrays of strings are accepted as input at fit and predict, like object arrays and DataFrames. (#1246)
• TabPFN-3.5 models work with the decoder-head readout in tabpfn-extensions. (#1274)
• TabPFN-3.5 gives correct predictions with inference_precision=torch.float16 on torch 2.5 to 2.8, where it returned near-uniform probabilities, and computes in float64 throughout when inference_precision=torch.float64 is set. (#1275)

Deprecated

• Passing an InferenceConfig object as inference_config is deprecated and now emits a FutureWarning. Unlike a dict, which overrides only the keys it names, an object replaces the checkpoint's config as a whole, so every field it does not set takes a class default rather than the checkpoint-specific value — which can silently degrade predictions. Pass a dict of just the settings you want to change; dataclasses.asdict(config) reproduces the old behavior for callers that do want a full replacement. (#1224)

Changelog

Sourced from tabpfn's changelog.

[9.0.0] - 2026-09-15

Breaking Changes

• eval_metric="roc_auc" combined with tune_decision_thresholds=True now raises instead of silently optimizing balanced accuracy. The threshold search scores thresholded labels, and the ROC AUC of a single operating point equals (TPR + TNR) / 2, so the two metrics produced identical thresholds. Pass eval_metric="balanced_accuracy" to keep the previous behaviour. (#1153)
• eval_metric="log_loss" combined with tune_decision_thresholds=True now raises instead of silently optimizing accuracy. The threshold search scores hard labels, and log loss over clipped hard labels is a constant multiple of the error rate. Use calibrate_temperature=True, which optimizes log loss directly. (#1239)
• TabPFN-3.5 is the default model: TabPFNClassifier() and TabPFNRegressor() now load ModelVersion.V3_5 unless a model_path or another version is given. The previous default, TabPFN-3, stays available through create_default_for_version(ModelVersion.V3). (#1273)

Added

• InferenceConfig.SOFTMAX_TEMPERATURE lets a checkpoint carry the softmax temperature it was trained for. softmax_temperature now defaults to "auto", meaning "use the checkpoint's value"; passing a float still overrides it for every model, though naming a temperature through both softmax_temperature and inference_config now raises. Existing checkpoints lack the field and fall back to its default of 0.9, the value the interface applied before, so nothing changes for them. Ensembling checkpoints that declare different temperatures now raises and asks for one to be passed explicitly. (#1220)
• InferenceConfig.N_ESTIMATORS lets a checkpoint declare how many estimators to run, and means exactly what the n_estimators argument means: an int is used as given and never raised for feature coverage, "auto" (the default) means 8 and may be raised. n_estimators="auto" now takes the checkpoint's value, so checkpoints without the field behave as before. An explicit n_estimators still wins; naming a count both ways, or ensembling checkpoints that disagree on it, raises. (#1230)
• inference_config={"TRANSFORM_DATES": True} expands a column holding a genuine datetime dtype (datetime64, tz-aware, or period) into calendar features: the year, the day of year, the seconds since the epoch, and the cyclical month, day and weekday pairs, plus the time of day when the column carries one. Off by default, in which case such a column is refused with a TabPFNValidationError naming it. Not run by the fine-tuning estimators. (#1236)
• inference_config={"TRANSFORM_TEXT": True} expands a text column into numeric features via skrub.StringEncoder: tf-idf over its character n-grams, reduced by a truncated SVD to at most TEXT_N_COMPONENTS features (30 by default). A column is text when it has pandas' string dtype, in any storage, or a pyarrow string dtype, more than MIN_CARDINALITY_FOR_TEXT distinct values, and is not declared categorical; an object or category column is never expanded. An unseen string is encoded by the character n-grams it shares with the training column; a missing value, or a string sharing none, becomes an all-zero row. Off by default, in which case nothing changes: such a column is ordinal-encoded as a high-cardinality category and fit() warns about it by name, as before, the warning now pointing at TRANSFORM_TEXT. Fitted estimators expose categorical_features_indices_, the declared categorical positions in the validated input after any date or text expansion. Not run by the fine-tuning estimators. (#1240)
• Add InferenceConfig.SAMPLE_SUBSAMPLING_METHOD to choose how rows are drawn per estimator when SUBSAMPLE_SAMPLES is set: "balanced", "stratified", or the new "majority_downsample", which groups rows by exact target value, keeps every row outside the single most frequent group, and fills the remaining budget from that majority group. This keeps the whole minority class for imbalanced classification and downsamples the repeated value for zero-inflated regression targets. Targets without a unique majority warn and fall back to the task's default sampling method. The default "auto" keeps the previous behavior (stratified for classification, balanced for regression). TabPFNClassifier.fit_with_differentiable_input now also uses stratified row subsampling, matching fit. (#1253)

Changed

• clean_data no longer casts a numeric input to float64; the preprocessing pipeline widens its own copy instead, cutting peak host RAM on a 1M x 2000 float32 fit + predict from 26.1 GB to 15.4 GB with identical outputs. (#1218)
• matplotlib is now a direct dependency instead of behind the optional viz extra (which has been removed), since skrub (a new direct dependency, used by upcoming work) already requires it transitively. (#1222)
• Widen the skrub dependency from the 0.10 series to 0.7 or newer. (#1231)
• The out-of-memory error now reports the feature count TabPFN actually ran on rather than the width of the frame passed to fit, which can be narrower when date expansion is on. (#1236)
• Dates are recognized only from a genuine datetime dtype, never by parsing string columns with pd.to_datetime(): a string column that merely looks like a date is read as a plain category or text. The never-consumed FeatureModality.DATE has been removed. (#1236)
• load_fitted_tabpfn_model and the load_from_fit_state class methods now default to device="auto" instead of device="cpu", so a GPU-fitted estimator no longer silently reloads onto CPU. Pass device="cpu" for the old behavior. (#1243)
• Declared categoricals are taken at face value: a string column listed in categorical_features_indices is categorical at any cardinality, never text, and a pandas category column now counts as declared (its position is merged into categorical_features_indices_). MIN_CARDINALITY_FOR_TEXT only applies to undeclared string columns. (#1245)
• Clarify estimator usage guidance for TabPFN-3 and later versions, covering dataset capacity, subsampling, and automatic handling of categorical and missing feature values without manual preprocessing. (#1247)
• Relicense the TabPFN code under standard Apache 2.0, removing the additional attribution requirement from the code license. Model weight licenses are unchanged: TabPFN-2 retains its existing Prior Labs License, and TabPFN-2.5, TabPFN-2.6, and TabPFN-3 remain non-commercial. (#1271)

Fixed

• • Fixed importance-based feature subsampling drawing every estimator's non-top features independently instead of from a shared round-robin pool, which could leave features unseen by any ensemble member even when the combined feature budget covered them all. Results change for fits using gini_feature_importance (directly or via "auto" on large, wide datasets). (#1141)
• Fix decision-threshold tuning assigning an arbitrary threshold to classes with no rows in the tuning holdout, which heavily boosted the class for eval_metric='roc_auc' and suppressed it for the other metrics. Such classes now receive a neutral threshold. (#1152)
• Cut TorchSquashingScaler's peak device memory from 13.0x its input to 0.4x for fit and 5.8x to 1.0x for transform, with byte-identical outputs, so large fits no longer exhaust VRAM. (#1223)
• Above 100k rows, importance-based feature subsampling gave every estimator its own feature ranking and its own sampling pool, so some features ended up in no ensemble member at all even when the combined feature budget covered them. A single shared ranking now restores full coverage, and replaces up to n_estimators LightGBM fits with one. Results change for gini_feature_importance fits (directly or via "auto") above 100k rows. (#1229)
• Prevent batched predictions from depending on other datasets' preprocessed feature widths. (#1233)
• get_embeddings now cleans categorical columns the way predict does, from the fitted feature schema rather than from the indices declared in categorical_features_indices. (#1236)
• TabPFNRegressor now reports the dataset sizes in its out-of-memory error. Only TabPFNClassifier passed them on, so the regressor's message omitted the "Your sizes" line entirely. (#1236)
• Fitting or predicting on a DataFrame with a genuine datetime column beside any other dtype, or with a timedelta64 or period column, no longer aborts with numpy's dtype-promotion error or Unknown dtype: period[D]. A datetime or period column is refused with a clear error, or expanded with TRANSFORM_DATES; a duration becomes its length in seconds. (#1236)
• Fixed TorchQuantileTransformer's sklearn-equivalence test failing on torch 2.14. torch 2.14 derives the sample index for a quantile in double precision, so the transformer's float32 quantile positions can land just short of an exact order statistic and get interpolated instead, shifting transformed values by up to 1.1e-4 where two neighbouring samples are close together. The test's tolerance now accommodates that; the transform still runs in float32, since computing the fit in float64 would cost performance in the pipeline for a difference this small. (#1238)
• Fixed saving a fitted estimator whose device is a torch.device (or a sequence of them), which raised TypeError: Object of type device is not JSON serializable. (#1243)
• numpy arrays of strings are accepted as input at fit and predict, like object arrays and DataFrames. (#1246)
• TabPFN-3.5 models work with the decoder-head readout in tabpfn-extensions. (#1274)
• TabPFN-3.5 gives correct predictions with inference_precision=torch.float16 on torch 2.5 to 2.8, where it returned near-uniform probabilities, and computes in float64 throughout when inference_precision=torch.float64 is set. (#1275)

Deprecated

... (truncated)

Commits

• c70b6ef Release v9.0.0 (#1283)
• 3018f39 Precision fixes in the multitask architecture: fp16 stays finite, fp64 stays ...
• 3420d0b CI: 60 minutes for the compatibility matrix (#1278)
• fb05cfe Add support for multitask parameter management in architecture classes (#1276)
• 2d94d79 Add the decoder attention readout to the multitask architecture (#1274)
• bcb6e2a Relicense TabPFN code under Apache 2.0 (#1271)
• 7065a12 Bump anthropics/claude-code-action from 1.0.202 to 1.0.217 (#1259)
• 75afafe Add a multitask architecture module and its model sources (#1256)
• 35e7045 Add majority_downsample row-subsampling mode via SAMPLE_SUBSAMPLING_METHOD (#...
• 428eb8a Take declared categoricals at face value: never TEXT, and fold in pandas cate...
• Additional commits viewable in compare view

Dependabot will resolve any conflicts with this PR as long as you don't alter it yourself. You can also trigger a rebase manually by commenting @dependabot rebase.

---

Dependabot commands and options

You can trigger Dependabot actions by commenting on this PR:

• @dependabot rebase will rebase this PR
• @dependabot recreate will recreate this PR, overwriting any edits that have been made to it
• @dependabot show <dependency name> ignore conditions will show all of the ignore conditions of the specified dependency
• @dependabot ignore this major version will close this PR and stop Dependabot creating any more for this major version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this minor version will close this PR and stop Dependabot creating any more for this minor version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this dependency will close this PR and stop Dependabot creating any more for this dependency (unless you reopen the PR or upgrade to it yourself)

[amrit110]

Automated fix applied and PR merged

The agentic fix loop successfully fixed this PR and merged it.

✓ Successfully fixed merge_only failures - Modified 0 files - Executed 94 agent actions - (58 info, 18 tool_call, 1 error, 15 tool_result, 2 reasoning)

View PR history on dashboard

AI Engineering Maintenance Bot