Fix: drop spurious label column in audiofolder/imagefolder datasets

services/worker/src/worker/job_runners/config/parquet_and_info.py

@@ -26,6 +26,7 @@
 from datasets.builder import DatasetBuilder
 from datasets.data_files import EmptyDatasetError as _EmptyDatasetError
 from datasets.download import StreamingDownloadManager
+from datasets.packaged_modules.audiofolder.audiofolder import AudioFolder as AudioFolderBuilder
 from datasets.packaged_modules.imagefolder.imagefolder import ImageFolder as ImageFolderBuilder
 from datasets.packaged_modules.parquet.parquet import Parquet as ParquetBuilder
 from datasets.packaged_modules.videofolder.videofolder import VideoFolder as VideoFolderBuilder
@@ -218,6 +219,15 @@ def is_video_builder(builder: DatasetBuilder) -> bool:
     return isinstance(builder, VideoFolderBuilder) or "Video(" in str(builder.info.features)
 
 
+def is_folder_based_builder(builder: DatasetBuilder) -> bool:
+    """Check if the builder is a folder-based builder (audiofolder/imagefolder).
+
+    These builders may incorrectly infer a 'label' column from directory structure
+    when directories are used for splits (train/test) rather than class labels.
+    """
+    return isinstance(builder, (AudioFolderBuilder, ImageFolderBuilder))
+
+
 def _is_too_big_from_hub(
     dataset_info: DatasetInfo,
     max_dataset_size_bytes: int,
@@ -1292,6 +1302,19 @@ def compute_config_parquet_and_info_response(
             token=hf_token,
             download_config=download_config,
         )
+        # For folder-based builders (audiofolder/imagefolder), reload with drop_labels=True
+        # to prevent spurious 'label' column when directories are used for splits (train/test)
+        # rather than class labels. See: https://github.com/huggingface/dataset-viewer/issues/3014
+        if is_folder_based_builder(builder):
+            logging.info(f"{dataset=} {config=} is a folder-based dataset, reloading with drop_labels=True")
+            builder = retry_load_dataset_builder(
+                path=dataset,
+                name=config,
+                revision=source_revision,
+                token=hf_token,
+                download_config=download_config,
+                drop_labels=True,
+            )
     except _EmptyDatasetError as err:
         raise EmptyDatasetError(f"{dataset=} is empty.", cause=err) from err
     except ValueError as err:

services/worker/tests/job_runners/config/test_parquet_and_info.py

@@ -42,6 +42,7 @@
     get_delete_operations,
     get_urlpaths_in_gen_kwargs,
     get_writer_batch_size_from_info,
+    is_folder_based_builder,
     limit_parquet_writes,
     list_generated_parquet_files,
     parse_repo_filename,
@@ -199,6 +200,31 @@ def test__is_too_big_from_datasets(
     )
 
 
+def test_is_folder_based_builder() -> None:
+    """Test is_folder_based_builder returns True for audiofolder/imagefolder builders."""
+    from datasets.packaged_modules.audiofolder.audiofolder import AudioFolder as AudioFolderBuilder
+    from datasets.packaged_modules.imagefolder.imagefolder import ImageFolder as ImageFolderBuilder
+    from datasets.packaged_modules.csv.csv import Csv as CsvBuilder
+    from unittest.mock import MagicMock
+
+    # Create mock builders
+    audio_builder = MagicMock(spec=AudioFolderBuilder)
+    audio_builder.__class__ = AudioFolderBuilder
+
+    image_builder = MagicMock(spec=ImageFolderBuilder)
+    image_builder.__class__ = ImageFolderBuilder
+
+    csv_builder = MagicMock(spec=CsvBuilder)
+    csv_builder.__class__ = CsvBuilder
+
+    # Test folder-based builders return True
+    assert is_folder_based_builder(audio_builder) is True
+    assert is_folder_based_builder(image_builder) is True
+
+    # Test non-folder-based builder returns False
+    assert is_folder_based_builder(csv_builder) is False
+
+
 def test_supported_if_big_parquet(
     app_config: AppConfig,
     get_job_runner: GetJobRunner,