Fixing #872 (#892)

Co-authored-by: Claude <noreply@anthropic.com>

Author: psinger-prior

changelog/892.fixed.md

@@ -0,0 +1 @@
+Fix `FinetunedTabPFNClassifier`/`FinetunedTabPFNRegressor` dropping pandas feature names from the final inference model. The raw training inputs are now retained so the fitted inference estimator records `feature_names_in_`, and calling `predict_proba`/`predict` with a DataFrame no longer triggers spurious sklearn feature-name warnings.

src/tabpfn/finetuning/finetuned_base.py

@@ -637,15 +637,18 @@ def _fit(  # noqa: C901,PLR0912
         self.finetuned_estimator_ = self._create_estimator(finetuning_estimator_config)
         self._setup_estimator()
 
-        X, y, _, _ = ensure_compatible_fit_inputs_sklearn(
-            X,
-            y,
-            estimator=self.finetuned_estimator_,
-            ensure_y_numeric=self._model_type == "regressor",
+        X_validated, y_validated, self.feature_names_in_, self.n_features_in_ = (
+            ensure_compatible_fit_inputs_sklearn(
+                X,
+                y,
+                estimator=self.finetuned_estimator_,
+                ensure_y_numeric=self._model_type == "regressor",
+            )
         )
-
         self.X_ = X
         self.y_ = y
+        X, y = X_validated, y_validated
+
         if X_val is not None and y_val is not None:
             X_train, y_train = X, y
             X_val, y_val, _, _ = ensure_compatible_fit_inputs_sklearn(