Raise clear error for problem_type="single_label_classification" with num_labels=1

src/transformers/configuration_utils.py

@@ -265,6 +265,19 @@ def __post_init__(self, **kwargs):
             # Keys are always strings in JSON so convert ids to int
             self.id2label = {int(key): value for key, value in self.id2label.items()}
 
+        # `problem_type="single_label_classification"` with `num_labels=1` is degenerate: applying
+        # cross-entropy to a single logit yields a constant zero loss. Reject this combination with a
+        # clear message pointing users to the intended setup (use `num_labels=2` for binary
+        # classification, or `problem_type="regression"` for a single-output regression head).
+        # See https://github.com/huggingface/transformers/issues/45479.
+        if self.problem_type == "single_label_classification" and self.num_labels == 1:
+            raise ValueError(
+                '`problem_type="single_label_classification"` requires `num_labels > 1`. With '
+                "`num_labels=1` the cross-entropy loss is degenerate and always zero. For binary "
+                'classification use `num_labels=2`, or use `problem_type="regression"` for a '
+                "single-output regression head."
+            )
+
         # BC for rotary embeddings. We will pop out legacy keys from kwargs and rename to new format
         if hasattr(self, "rope_parameters"):
             kwargs = self.convert_rope_params_to_dict(**kwargs)

tests/utils/test_configuration_utils.py

@@ -261,6 +261,20 @@ def test_loading_config_do_not_raise_future_warnings(self):
             warnings.simplefilter("error")
             PreTrainedConfig.from_pretrained("bert-base-uncased")
 
+    def test_single_label_classification_requires_more_than_one_label(self):
+        """Regression test for https://github.com/huggingface/transformers/issues/45479.
+
+        `problem_type="single_label_classification"` with `num_labels=1` used to silently produce a
+        degenerate zero cross-entropy loss. It must now raise a clear error at config construction.
+        """
+        with self.assertRaises(ValueError):
+            BertConfig(num_labels=1, problem_type="single_label_classification")
+
+        # Valid combinations must still work.
+        BertConfig(num_labels=2, problem_type="single_label_classification")
+        BertConfig(num_labels=1, problem_type="regression")
+        BertConfig(num_labels=1)  # problem_type left unset is fine; it is inferred at forward time.
+
     def test_get_text_config(self):
         """Tests the `get_text_config` method."""
         # 1. model with only text input -> returns the original config instance