AbstentionBench (Parquet snapshot, private mirror)

This is a Parquet snapshot of facebook/AbstentionBench. It loads with current datasets (4.x) and needs no trust_remote_code:

from datasets import load_dataset
ds = load_dataset("ymerkli/AbstentionBench")            # DatasetDict, one split per source
falseqa = load_dataset("ymerkli/AbstentionBench", split="falseqa")

• Snapshot date: 2026-10-21
• Upstream script commit: af06080e2cbfff73f02a7c33135e5567d6f7ce6a (AbstentionBench.py + data.py)
• Built with: datasets==3.6.0, using upstream source loaders and upstream subsampling-indices.json / UMWP_indices_answerable.json

Structure

Each split is one source dataset, named like the upstream split. Fields:

• question (str), reference_answers (list[str] or null), should_abstain (bool), metadata_json (JSON-encoded str): unchanged from upstream.
• source_dataset (str): added in this mirror. It equals the upstream split name the row came from.

To get the upstream "all splits" view, use datasets.concatenate_datasets(list(ds.values())).

Differences from upstream

• gpqa_abstain is excluded (80 rows upstream). Idavidrein/gpqa is gated, and the account that built this snapshot did not have access. No GPQA data was regenerated or substituted.
• QAQA extraction patch. The only code change in the local loader copy makes the QAQA tarball extract with tarfile.extractall(..., filter="data"), which blocks path traversal. It does not affect the extracted data.
• Builds ran per source with the same constructor arguments, subsampling indices and order as upstream download_and_prepare, so one gated source would not abort the whole build.
• Upstream behaviour kept as-is (not "fixed"):
  • known_unknown_questions is defined twice in the upstream script. The second definition, without category_map_path, wins, so KUQ_category is null and kuq_new_categories.csv is not applied.
  • The paper says KUQ's counterfactual category is excluded, but the 3500-row subsample is drawn from all of KUQ.
  • musique is not subsampled. The NQDataset indices in subsampling-indices.json are keyed by class name and do not apply to MusiqueDataset.

Row counts

All 19 included splits match the expected upstream sizes: 3500-row caps from subsampling-indices.json, plus the upstream repo's tests/test_datasets.py and recounts of the sources. Expected upstream total with GPQA: 35,935.

split	rows	should_abstain=True	should_abstain=False
alcuna	3500	1732	1768
bbq	3500	1778	1722
big_bench_disambiguate	250	75	175
big_bench_known_unknowns	46	23	23
coconot	1380	1001	379
falseqa	1374	687	687
gsm8k_abstain	2426	1213	1213
known_unknown_questions	3500	1760	1740
mediq	2818	1409	1409
mmlu_history_abstain	78	39	39
mmlu_math_abstain	266	133	133
moral_choice	1367	680	687
musique	3266	1316	1950
qaqa	570	285	285
qasper	1287	79	1208
situated_qa	635	129	506
squad2	3500	1771	1729
umwp	3500	1750	1750
world_sense	2592	864	1728
total	35855

License

CC BY-NC 4.0 (Meta's contribution, as on the upstream card). Upstream note: "Third party content pulled from other locations are subject to its own licenses and you may have other legal obligations or restrictions that govern your use of that content." Each source dataset keeps its own license.

Citation

@misc{kirichenko2025abstentionbenchreasoningllmsfail,
      title={AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions},
      author={Polina Kirichenko and Mark Ibrahim and Kamalika Chaudhuri and Samuel J. Bell},
      year={2025},
      eprint={2506.09038},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2506.09038},
}