add back classes_ attribute

tabpfn_client/estimator.py

@@ -3,8 +3,8 @@
 
 from typing import Optional, Literal, Dict, Union
 import logging
-
 import numpy as np
+import pandas as pd
 from tabpfn_client.config import init
 from sklearn.base import BaseEstimator, ClassifierMixin, RegressorMixin
 from sklearn.utils.validation import check_is_fitted
@@ -134,6 +134,7 @@ def fit(self, X, y):
         init()
 
         validate_data_size(X, y)
+        self._validate_targets_and_classes(y)
         _check_paper_version(self.paper_version, X)
 
         estimator_param = self.get_params()
@@ -192,6 +193,17 @@ def _predict(self, X, output_type):
         )
         return res
 
+    def _validate_targets_and_classes(self, y) -> np.ndarray:
+        from sklearn.utils import column_or_1d
+        from sklearn.utils.multiclass import check_classification_targets
+
+        y_ = column_or_1d(y, warn=True)
+        check_classification_targets(y)
+        # Get classes and encode before type conversion to guarantee correct class labels.
+        not_nan_mask = ~pd.isnull(y)
+        # TODO: should pass this from the server
+        self.classes_ = np.unique(y_[not_nan_mask])
+
 
 class TabPFNRegressor(BaseEstimator, RegressorMixin, TabPFNModelSelection):
     _AVAILABLE_MODELS = [

tabpfn_client/tests/unit/test_tabpfn_classifier.py

@@ -434,6 +434,46 @@ def test_predict_params_output_type(self):
             predict_params = mock_predict.call_args[1]["predict_params"]
             self.assertEqual(predict_params, {"output_type": "probas"})
 
+    def test_string_label_predictions(self):
+        """Test that string labels in y are preserved in predictions."""
+        X = np.random.rand(10, 5)
+        test_X = np.random.rand(5, 5)
+
+        # Test with pure string labels
+        y_str = np.array(["cat", "dog"] * 5)
+        clf_str = TabPFNClassifier()
+
+        # Mock fit and predict
+        with patch.object(InferenceClient, "fit") as mock_fit:
+            mock_fit.return_value = "dummy_uid"
+            clf_str.fit(X, y_str)
+
+            with patch.object(InferenceClient, "predict") as mock_predict:
+                mock_predict.return_value = np.array(["cat", "dog"] * 5)
+                y_pred = clf_str.predict(test_X)
+
+                self.assertTrue(np.all(np.isin(y_pred, ["cat", "dog"])))
+                self.assertTrue(
+                    y_pred.dtype.kind in {"U", "O"}, "Predictions should be string type"
+                )
+
+        # Test with numeric-like strings
+        y_num_str = np.array(["0", "1"] * 5)
+        clf_num_str = TabPFNClassifier()
+
+        with patch.object(InferenceClient, "fit") as mock_fit:
+            mock_fit.return_value = "dummy_uid"
+            clf_num_str.fit(X, y_num_str)
+
+            with patch.object(InferenceClient, "predict") as mock_predict:
+                mock_predict.return_value = np.array(["0", "1"] * 5)
+                y_pred = clf_num_str.predict(test_X)
+
+                self.assertTrue(np.all(np.isin(y_pred, ["0", "1"])))
+                self.assertTrue(
+                    y_pred.dtype.kind in {"U", "O"}, "Predictions should be string type"
+                )
+
 
 class TestTabPFNModelSelection(unittest.TestCase):
     def setUp(self):