Raise early for invalid revision in load_dataset (#7929)

* Add test for checking invalid revision handling

* Fix load_dataset ignoring invalid revision when cache exists

Author: Scott-Simmons

src/datasets/load.py

@@ -979,10 +979,6 @@ def dataset_module_factory(
                 elif e.response.status_code == 403:
                     message += f" Visit the dataset page at https://huggingface.co/datasets/{path} to ask for access."
                 raise DatasetNotFoundError(message) from e
-            except RevisionNotFoundError as e:
-                raise DatasetNotFoundError(
-                    f"Revision '{revision}' doesn't exist for dataset '{path}' on the Hub."
-                ) from e
             except RepositoryNotFoundError as e:
                 raise DatasetNotFoundError(f"Dataset '{path}' doesn't exist on the Hub or cannot be accessed.") from e
             try:
@@ -1016,10 +1012,8 @@ def dataset_module_factory(
                 elif e.response.status_code == 403:
                     message += f" Visit the dataset page at https://huggingface.co/datasets/{path} to ask for access."
                 raise DatasetNotFoundError(message) from e
-            except RevisionNotFoundError as e:
-                raise DatasetNotFoundError(
-                    f"Revision '{revision}' doesn't exist for dataset '{path}' on the Hub."
-                ) from e
+        except RevisionNotFoundError as e:
+            raise DatasetNotFoundError(f"Revision '{revision}' doesn't exist for dataset '{path}' on the Hub.") from e
         except Exception as e1:
             # All the attempts failed, before raising the error we should check if the module is already cached
             try:

tests/test_load.py

@@ -763,6 +763,18 @@ def test_load_dataset_from_hub(self):
                         str(context.exception),
                     )
 
+    @pytest.mark.integration
+    def test_load_dataset_invalid_revision_with_cache(self):
+        repo_id = SAMPLE_DATASET_IDENTIFIER2
+        builder = load_dataset_builder(repo_id, cache_dir=self.cache_dir)
+        builder.download_and_prepare()
+        with self.assertRaises(DatasetNotFoundError) as context:
+            datasets.load_dataset(repo_id, revision="invalid_revision", cache_dir=self.cache_dir)
+        self.assertIn(
+            "Revision 'invalid_revision' doesn't exist for dataset",
+            str(context.exception),
+        )
+
     def test_load_dataset_namespace(self):
         with self.assertRaises(DatasetNotFoundError) as context:
             datasets.load_dataset("hf-internal-testing/_dummy")