Breaking Changes
fitnow acceptssample_weight, so scikit-learn meta-estimators pass weights to it, andfitraises for models withoutSAMPLE_WEIGHT_MODE.BaggingClassifier(TabPFNClassifier())andBaggingRegressor(TabPFNRegressor())raise infitwith anybootstrapandmax_samples, because they pass their row counts as weights instead of fitting on a row subset.CalibratedClassifierCV(TabPFNClassifier()).fit(X, y, sample_weight=w)raises instead of warning that the estimator ignores the weights. No released model supports sample weights yet. (#1415)- A missing value in a
group_colcolumn now raises infitandpredict. Before, such a row was treated as belonging to no known group. (#1418)
Added
- Add the
log_quantile_gap,log_fixed_atomandlog1p_targetregression target transforms for nonnegative targets.log_quantile_gapandlog_fixed_atomdecode a range to exactly 0 for zero-inflated targets.log1p_targetencodes log1p(y) of the raw target and decodes negative codes to 0. (#1409) - Add
InferenceConfig.REGRESSION_READOUT. With"nonnegative_and_capped", the regression outputs of a nonnegative target are read from the ensemble's averaged distribution of max(Y, 0) given Y <= y_max, with y_max the largest training target. (#1410) - Add
RegressorTuningConfig.calibrate_mean. It multiplies the predicted mean by a constant fitted on the tuning holdout rows. Its onlymean_calibration_mode,"tweedie", picks the constant that minimizes the Tweedie deviance (powermean_calibration_power, default 1.5). (#1411) RegressorTuningConfig.calibrate_meanweights each tuning holdout row bysample_weight * label_reliability. For an exposure, pass the rate as the target and the exposure assample_weight. (#1412)TabPFNClassifier.fitandTabPFNRegressor.fittake experimentalsample_weightandlabel_reliabilityarguments, applied by models whose inference config setsSAMPLE_WEIGHT_MODE. For every other modelfitraises when they are given. (#1415)- Two
inference_configflags add transductive modes forgroup_coldata:APPEND_GROUP_AGGREGATESappends per-group aggregates of the features, andPOOL_GROUP_PREDICTIONSgives the rows of a group their mean or geometric-mean prediction when every group has one label. A missing value in agroup_colcolumn is now refused. (#1418) sample_weightandlabel_reliabilitycan be combined withtuning_config: each tuning fold is fit with the weights of its rows, and a holdout row counts in the tuned metric in proportion tosample_weight * label_reliability. (#1428)- Add
InferenceConfig.AVERAGE_BEFORE_SOFTMAX, so a checkpoint can set whether the ensemble averages logits or probabilities.average_before_softmaxnow defaults to"auto", which uses the checkpoint's value. (#1432)
Changed
- Lower peak host and GPU memory when fitting and predicting on large tables: preprocessing works in place by blocks, a float32 input stays float32 where the values are not changed, and the model input can stay in host memory for a model whose
forwardtakesx_host. Addtabpfn.preprocessing.clean.overwrite_fit_input()to letfitwrite into an array the caller will not read again. (#1422) - With
ENABLE_GPU_PREPROCESSING, therobustpreset runs on the GPU in float32. (#1424) - With
APPEND_GROUP_AGGREGATES, the fit selects the per-group aggregates faster: onn_preprocessing_jobsthreads, and for numeric columns on the GPU whenENABLE_GPU_PREPROCESSINGis set and the estimator runs on a CUDA device. The default of one job on CPU is unchanged. (#1429)
Fixed
TabPFNRegressor.predictwithfit_mode="fit_with_cache"now predicts large test sets in blocks of rows, so peak memory no longer grows with the number of test rows. The block size is set byTABPFN_MAX_REGRESSOR_PREDICT_ROWS(default 65,536; 0 disables it). (#1417)