diff --git a/lightautoml/automl/base.py b/lightautoml/automl/base.py
index c29fcba3..69987c23 100644
--- a/lightautoml/automl/base.py
+++ b/lightautoml/automl/base.py
@@ -185,6 +185,20 @@ def fit_predict(
         self.timer.start()
         train_dataset = self.reader.fit_read(train_data, train_features, roles)
 
+        # Saving class mapping
+        if self.reader.task.name == "binary":
+            self.classes_ = [1]
+        elif self.reader.task.name == "multi:reg":
+            self.classes_ = roles["target"]
+        elif self.reader.task.name == "reg":
+            self.classes_ = [roles["target"]]
+        else:
+            self.classes_ = (
+                sorted(self.reader.class_mapping, key=self.reader.class_mapping.get, reverse=False)
+                if self.reader.class_mapping
+                else None
+            )
+
         assert (
             len(self._levels) <= 1 or train_dataset.folds is not None
         ), "Not possible to fit more than 1 level without cv folds"
@@ -259,7 +273,7 @@ def fit_predict(
             else:
                 break
 
-        blended_prediction, last_pipes = self.blender.fit_predict(level_predictions, pipes)
+        blended_prediction, last_pipes = self.blender.fit_predict(level_predictions, pipes, self.classes_)
         self.levels.append(last_pipes)
 
         self.reader.upd_used_features(remove=list(set(self.reader.used_features) - set(self.collect_used_feats())))
diff --git a/lightautoml/automl/blend.py b/lightautoml/automl/blend.py
index 7060270b..d1ba9706 100644
--- a/lightautoml/automl/blend.py
+++ b/lightautoml/automl/blend.py
@@ -42,7 +42,7 @@ def outp_dim(self) -> int:  # noqa: D102
         return self._outp_dim
 
     def fit_predict(
-        self, predictions: Sequence[LAMLDataset], pipes: Sequence[MLPipeline]
+        self, predictions: Sequence[LAMLDataset], pipes: Sequence[MLPipeline], class_mapping: dict
     ) -> Tuple[LAMLDataset, Sequence[MLPipeline]]:
         """Wraps custom ``._fit_predict`` methods of blenders.
 
@@ -63,7 +63,7 @@ def fit_predict(
             self._bypass = True
             return predictions[0], pipes
 
-        return self._fit_predict(predictions, pipes)
+        return self._fit_predict(predictions, pipes, class_mapping)
 
     def _fit_predict(
         self, predictions: Sequence[LAMLDataset], pipes: Sequence[MLPipeline]
@@ -134,7 +134,7 @@ def split_models(self, predictions: Sequence[LAMLDataset]) -> Tuple[Sequence[LAM
 
         return splitted_preds, model_idx, pipe_idx
 
-    def _set_metadata(self, predictions: Sequence[LAMLDataset], pipes: Sequence[MLPipeline]):
+    def _set_metadata(self, predictions: Sequence[LAMLDataset], pipes: Sequence[MLPipeline], class_mapping: dict):
 
         pred0 = predictions[0]
         pipe0 = pipes[0]
@@ -143,6 +143,8 @@ def _set_metadata(self, predictions: Sequence[LAMLDataset], pipes: Sequence[MLPi
         self._outp_prob = pred0.task.name in ["binary", "multiclass"]
         self._score = predictions[0].task.get_dataset_metric()
 
+        self._class_mapping = class_mapping
+
     def score(self, dataset: LAMLDataset) -> float:
         """Score metric for blender.
 
@@ -321,7 +323,7 @@ def _get_weighted_pred(self, splitted_preds: Sequence[NumpyDataset], wts: Option
         outp = splitted_preds[0].empty()
         outp.set_data(
             weighted_pred,
-            ["WeightedBlend_{0}".format(x) for x in range(weighted_pred.shape[1])],
+            self._class_mapping if self._class_mapping else list(range(weighted_pred.shape[1])),
             NumericRole(np.float32, prob=self._outp_prob),
         )
 
@@ -436,7 +438,7 @@ def _prune_pipe(
         return new_pipes, wts
 
     def _fit_predict(
-        self, predictions: Sequence[NumpyDataset], pipes: Sequence[MLPipeline]
+        self, predictions: Sequence[NumpyDataset], pipes: Sequence[MLPipeline], class_mapping: dict
     ) -> Tuple[NumpyDataset, Sequence[MLPipeline]]:
         """Perform coordinate descent.
 
@@ -451,7 +453,7 @@ def _fit_predict(
             Dataset and MLPipeline.
 
         """
-        self._set_metadata(predictions, pipes)
+        self._set_metadata(predictions, pipes, class_mapping)
         splitted_preds, _, pipe_idx = cast(List[NumpyDataset], self.split_models(predictions))
 
         wts = self._optimize(splitted_preds)