You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Python: [Bug]: GAIA runner refuses local parquet data and asks for HF_TOKEN #9337
_load_gaia_local prefers parquet files (comment calls it the new format, validation split first) but GAIA._ensure_data only treats the data dir as ready when it contains metadata.jsonl. So a data_dir with only parquet files is ignored and the runner demands HF_TOKEN and tries a fresh download even though the local data is fully usable.
Expected Behavior
_ensure_data should accept the parquet layout the loader itself supports.
Steps to Reproduce
Create a data dir with 2023/validation/metadata-00000-of-00001.parquet holding two task rows.
_load_gaia_local(dir) returns both tasks.
GAIA(data_dir=dir)._ensure_data() raises RuntimeError asking for HF_TOKEN.
Minimal Reproduction
importtempfilefrompathlibimportPathimportpyarrowaspaimportpyarrow.parquetaspqfromagent_framework_lab_gaia.gaiaimport_load_gaia_local, GAIAtmp=Path(tempfile.mkdtemp())
val=tmp/"2023"/"validation"val.mkdir(parents=True)
rows= [
{"task_id": "v1", "Question": "What is 2+2?", "Final answer": "4", "Level": 1, "file_name": "", "Annotator Metadata": "{}"},
{"task_id": "v2", "Question": "Capital of France?", "Final answer": "Paris", "Level": 1, "file_name": "", "Annotator Metadata": "{}"},
]
pq.write_table(pa.Table.from_pylist(rows), val/"metadata-00000-of-00001.parquet")
assertlen(_load_gaia_local(tmp)) ==2# loader is fine with parquetGAIA(data_dir=str(tmp))._ensure_data() # RuntimeError: HF_TOKEN ... required
Error Messages and Stack Traces
RuntimeError: HF_TOKEN environment variable or hf_token parameter is required to access the GAIA dataset. Please set your Hugging Face token with access to gaia-benchmark/GAIA.
Looks like _ensure_data was written for the old jsonl layout and never updated when parquet loading was added. Accepting metadata*.parquet in the same existence check would fix it. Happy to send a PR.
Acknowledgements
I searched existing issues and did not find a duplicate.
The fix for this issue is resubmitted as #9609 (replacing the closed #9340, which was closed with a request to address the outstanding review feedback: _ensure_data() accepted a directory whenever any metadata.jsonl existed, without checking incomplete parquet shards in the same directory). The new PR contains the original parquet-only dir fix plus that feedback: any mixed layout with metadata parquet present now requires the complete parquet snapshot, with a regression test for the mixed layout.
Activity
ManoharPaturi commentedon Oct 22, 2026
The fix for this issue is resubmitted as #9609 (replacing the closed #9340, which was closed with a request to address the outstanding review feedback:
_ensure_data()accepted a directory whenever anymetadata.jsonlexisted, without checking incomplete parquet shards in the same directory). The new PR contains the original parquet-only dir fix plus that feedback: any mixed layout with metadata parquet present now requires the complete parquet snapshot, with a regression test for the mixed layout.