GAIA: a benchmark for General AI Assistants
Paper • 2311.12983 • Published • 252
Datasets:
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This repository contains GAIA validation questions, reference answers, linked variants, and attachments. Share GAIA-derived material only through gated or private Hugging Face repositories, and do not make validation or test content crawlable. See LICENSE for the source conditions and attribution.
Log in or Sign Up to review the conditions and access this dataset content.
REUSE-QA GAIA
This is the GAIA component of REUSE-QA, a benchmark for studying how tool-using agents reuse work across related questions. It contains 318 questions in 139 families: 139 originals and 179 linked variants from GAIA's 2023 validation Levels 1 and 2.
The full REUSE-QA benchmark has 7,797 questions in 2,627 families across two repositories. RECAP-data/reuse-qa contains the 7,479 FinanceBench and SQL-API-Bench questions with ungated access. This separate GAIA repository contains the remaining 318 questions under GAIA's access conditions.
Contents and access
The default configuration has two splits, selecting GAIA levels: gaia_l1 contains 146 questions and gaia_l2 contains 172. The composition is as follows:
| Source | Questions | Originals | P | ER | CA |
|---|---|---|---|---|---|
| GAIA | 318 | 139 | 139 | 17 | 23 |
P means paraphrase, ER entity replacement, and CA condition augmentation. All 31 attachments referenced by 62 records are included under 2023/validation/ (8.27 MB).
Sign in on this dataset's Hugging Face page and accept the source conditions for automatic access approval. Then authenticate locally:
pip install datasets huggingface_hub
from datasets import load_dataset
from huggingface_hub import login
login() # Use an account that has accepted this repository's access conditions.
questions = load_dataset(
"RECAP-data/reuse-qa-gaia", "default", split="gaia_l1", token=True
)
For sequential iteration without loading the complete split, add streaming=True to load_dataset. The Hugging Face streaming guide describes this option.
Record format
Each record has eight top-level fields:
| Field | Meaning |
|---|---|
id |
Unique record identifier. |
parent_id |
Original question's ID for a variant; null for an original. |
dataset |
gaia_l1 or gaia_l2. |
variant_type |
original, paraphrase, entity_replacement, or condition_augmentation. |
question |
Question presented to the agent. |
answer |
Reference answer as a string; keep it out of the agent's inputs. |
answer_type |
text or numeric. |
context |
Task inputs, including the GAIA level and any attachment reference. |
The shared REUSE-QA context schema has eight fields: db_name, company, doc_name, evidence, constraint_type, file_name, file_path, and level. GAIA uses level and, where applicable, file_name and file_path. The other fields are null, except evidence, which is an empty list. The source compilation's top-level metadata annotations are omitted.
A variant's parent is in the same split. Paraphrases retain the original reference answer. Entity replacements change a query-specific input; condition variants add or revise a constraint. Use variant_type and parent_id to identify these roles, since IDs retain older construction markers.
Attachments and tools
Download a repository snapshot and resolve context.file_path relative to its root:
from pathlib import Path
from datasets import load_dataset
from huggingface_hub import snapshot_download
repo_root = Path(snapshot_download(
"RECAP-data/reuse-qa-gaia", repo_type="dataset", token=True
))
questions = load_dataset(
"RECAP-data/reuse-qa-gaia", "default", split="gaia_l1", token=True
)
record = next(row for row in questions if row["context"]["file_path"])
attachment = repo_root / record["context"]["file_path"]
The attachments are unchanged copies from GAIA revision 682dd723ee1e1697e00360edccf2366dc8418dd9. Execution also requires tools for web search, page access, computation, and attachment handling. Bundled files do not freeze the web sources or tool environment.
Source conditions and citation
GAIA's source conditions at the pinned revision limit resharing to gated or private Hugging Face repositories and prohibit making validation or test content crawlable. Preserve these conditions when further sharing GAIA-derived material. LICENSE records the source conditions and attribution.
Please cite the accompanying manuscript, RECAP: Structured Plan Reuse for Efficient Tool-Using Agents, and the source benchmark:
Gregoire Mialon, Clementine Fourrier, Craig Swift, Thomas Wolf, Yann LeCun, and Thomas Scialom. GAIA: A Benchmark for General AI Assistants. ICLR 2024. Paper · Dataset.
- Downloads last month
- 26