BEANS-Next

Audio files live under audio/. Metadata is in metadata.parquet with a file_name column (repo-relative paths) so the Hugging Face Dataset Viewer can play clips—only file_name uses that convention; tier 4 rows instead use context_audio_paths (list) and query_audio_path so the viewer is not confused by several *_file_name-like columns. Column task is the benchmark task id (same strings as the old subset column in other layouts). Column tier is an integer in {1,2,3,4} (legacy Parquet may still use directory strings; loaders normalize both). Example: dataset["test"].filter(lambda x: x["task"] == "crow-description").