GAIA: a benchmark for General AI Assistants
Paper • 2311.12983 • Published • 249
Acknowledge GAIA's terms to access GAIA-Verified
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
GAIA-Verified is derived from GAIA and is redistributed here under GAIA's own condition that the validation set not be reshared outside a gated or private repository on the HF hub. By accessing it you agree to the same terms you would accept for GAIA, including not resharing it in a crawlable format.
Log in or Sign Up to review the conditions and access this dataset content.
GAIA-Verified
A corrected subset of GAIA's validation split: 147 of the original 165 tasks.
27 of GAIA's 165 validation tasks (16.4%) are defective. 18 could not be salvaged and were removed; 9 had a gold answer that is simply wrong and were corrected. Questions were never rewritten and the scorer was never patched.
| Defective | 27 / 165 (16.4%) |
| Removed | 18 |
| Golds corrected | 9 |
| GAIA-Verified | 147 tasks |
Every verdict, with its evidence, is in the audit: https://github.com/anirudhvenu/gaia-verified
What changed
- Source has drifted or died. Gold was right when annotated; the web has since moved — databases recomputed, records backfilled, pages deleted, affordances retired.
- Question does not yield the gold. Read the question literally and you get a different answer than the annotator recorded.
- Gold was wrong when written. Plain annotator error, provable then and now.
- Gold breaks GAIA's own answer-format rule. GAIA's prompt forbids articles; two golds require one, and the scorer never strips them, so a compliant model could not pass.
Staleness and question-wording dominate; outright annotator error is the minority. GAIA is not sloppily built — it is built on a web that has moved.
Use it
from datasets import load_dataset
ds = load_dataset("anivenu/gaia-verified", "2023_all", split="validation") # 147 tasks
The configs mirror GAIA's (2023_all, 2023_level1, 2023_level2, 2023_level3), so an
existing harness works by swapping the dataset id.
Schema
Identical to GAIA, plus one column. The 9 corrected rows carry
gaia_verified_original_answer, holding the value that Final answer replaced. Every other row
is byte-identical to GAIA.
Provenance
Derived from gaia-benchmark/GAIA, 2023/validation, revision 682dd723. Audited July 2026.
The corrections are reproducible from the errata alone — see the repo above — so you can rebuild
this dataset from your own copy of GAIA rather than trusting this one.
Citation
Cite GAIA. This is an errata layer over their work, not a new benchmark.
@misc{mialon2023gaia,
title={GAIA: a benchmark for General AI Assistants},
author={Mialon, Gr\'egoire and Fourrier, Cl\'ementine and Swift, Craig and Wolf, Thomas
and LeCun, Yann and Scialom, Thomas},
year={2023}, eprint={2311.12983}, archivePrefix={arXiv}
}
- Downloads last month
- -