VideoGAIA-271

VideoGAIA-271 is a human-verified benchmark for agentic video understanding. It contains 271 open-ended tasks that require a model to combine visual-temporal evidence with targeted tool use and external knowledge.

Files

videogaia_data/
├── README.md
├── videogaia_271.jsonl
└── videos/
    └── 271 MP4 videos

All paths stored in the JSONL are relative to this directory, so the package can be moved without rewriting metadata.

Statistics

Category	Tasks
Culture	58
Geography	63
Daily Life	61
History	44
Technology	39
Economics	6
Total	271

Difficulty	Tasks
L1	62
L2	209

L1 contains the relatively more accessible tasks, while L2 emphasizes harder multi-step evidence acquisition, cross-modal reasoning, and verification.

JSONL Schema

Each line is one JSON object with the following fields:

Field	Type	Description
sample_id	string	Stable sample identifier
video.local_path	string	Relative path to the MP4 file
video.duration_sec	number	Video duration in seconds
question_en	string	English open-ended question used for evaluation
answer	string	Human-verified reference answer
category	string	One of the six VideoGAIA categories
difficulty	string	L1 or L2

Example:

{"sample_id":"vdr_example","video":{"local_path":"videos/vdr_example.mp4","duration_sec":21.2},"question_en":"...","answer":"...","category":"Technology","difficulty":"L1"}

Loading

import json
from pathlib import Path

root = Path("videogaia_data")
with (root / "videogaia_271.jsonl").open(encoding="utf-8") as handle:
    samples = [json.loads(line) for line in handle if line.strip()]

first_video = root / samples[0]["video"]["local_path"]
print(samples[0]["question_en"], first_video)

The companion videogaia repository contains the official agent-loop inference and evaluation harness.

Citation

Citation metadata and the final redistribution terms should be added alongside the public VideoGAIA paper release.