VibeVideoBench (dataset)

Tasks, rubrics and quizzes of VibeVideoBench, a benchmark for AI agents that produce finished videos. Each task is a production brief with source material; an agent works in an isolated sandbox for up to one hour and delivers an MP4 plus the editable project behind it. The video is then checked by programs and graded by AI judges against the task's rubric.

Paper: https://arxiv.org/abs/2610.35864 · Project page: https://vibevideobench.github.io

This repository holds the data only. The evaluation pipeline — sandbox image, agent runner, judges and report — is on GitHub: vibevideobench/vibevideobench.

tasks	239 — 209 English, 30 Chinese; 205 main tasks and 34 education tasks (videos that explain a paper or a lesson)
rubric checks	12,985 binary checks (pass / fail / unknown), each answerable from the finished video
quality dimensions	5 per task (image quality, visual design, clarity, timing, audio), scored 1–5 against task-specific anchors, with 4,780 observation points
program checks	1,962 L1 checks on the delivered file (decode, size, frame rate, duration, audio, speech, language ...)
quiz	404 multiple-choice questions on the 34 education tasks, grounded in the source material

Layout

tasks/<ID>/
  prompt.txt          the brief the agent receives
  task.json           the same task, structured: delivery spec (size, fps, duration, audio_required, ...), audio policy
  inputs/             source material (images, audio, data, fonts, code) in the task's language
tasks/index.json      task list: id, title, language
rubrics/<ID>.json     core goals, the rubric checks (task_rubric), the five quality dimensions (quality_rubric) and
                      the L1 program checks (program_check_proposals)
tags/<ID>.json        modality tag of every check and observation point: visual / audio / audio_visual / artifact;
                      it decides which judges see the item
quiz/<ID>.json        education tasks: questions, options and the answer
EXTERNAL_INPUTS.md    inputs you download yourself (below); external.json is the same list for programs
NOTICE.md             credits and licenses of the third-party inputs that are included
LICENSE               CC BY-NC-SA 4.0
fetch_external.py     optional helper that downloads and checks them

Each task's inputs are in the task's own language. Where another language is itself the subject (glyphs or signage being depicted, bilingual fields) or the file is original evidence (photos, scans, official PDFs), it is kept as is.

Inputs you download yourself

62 inputs of 23 tasks are third-party material we may not redistribute: arXiv papers (PA01–PA09), Adobe font documentation (R0003), Apple Newsroom press images (R0011), an FAO report and a UN web page (D0130), FIFA web pages (N0001), excerpts of an ESA video (D0248) and of a NASA video (W0160), photos of identifiable people (R0001, R0005, R0007, R0013), Pexels stock photos (W0136, W0137) and two brand marks (W0010, W0342). Put them in place before you evaluate anything. EXTERNAL_INPUTS.md gives, for each file, the path it goes to, where to get it, the sha256 of the copy the benchmark was run with and, for the video excerpts, a ready ffmpeg command. python3 fetch_external.py fetches everything in one go (it uses your http(s)_proxy) and python3 fetch_external.py --check lists what is still missing.

Downloads are checked against the benchmark's copies: byte for byte for plain files; duration, frame size, frame rate and audio streams for the re-encoded video excerpts; by their text for web pages, whose markup changes between captures (they come from the Internet Archive, or the live page when no capture is available).

Use

# the pipeline expects the dataset in its data/ folder
git clone https://github.com/vibevideobench/vibevideobench && cd vibevideobench
huggingface-cli download diaomuxi/vibevideobench --repo-type dataset --local-dir data
python3 data/fetch_external.py

Then follow the pipeline's README. Without the pipeline, each task is plain files: prompt.txt is the brief and rubrics/<ID>.json is what the judges check.

License and attribution

The tasks, briefs, rubrics, tags and quizzes are licensed under CC BY-NC-SA 4.0: free to use, share and adapt for non-commercial purposes, with attribution, and adaptations under the same license. Third-party material among the task inputs keeps its own license — public domain, CC0, CC BY / CC BY-SA, ODbL, MIT, the SIL Open Font License, GPL with the font exception — and NOTICE.md lists each with its credit line. Font and software license texts are included in each task's inputs.

Citation

@misc{diao2026vibevideobenchevaluatingmultimodalagents,
      title={VibeVideoBench: Evaluating Multimodal Agents on Code-Driven Video Generation},
      author={Muxi Diao and Wenbin Xu and Lele Yang and Xuyang Ma and Yuzhe Chen and Jiankang Chen and Kaiyu Jiang and Changyi Liu and Haoyu Zhang and Hongyi Fu and Jianxiong Wang and Tianke Zhang and Haonan Fan and Yingxin Li and Bingyi Li and Xu Wang and Xuanyu Zheng and Bin Wen and Tingting Gao and Han Li and Wentao Zhang and Weiran Xu and Kongming Liang and Zhanyu Ma},
      year={2026},
      eprint={2610.35864},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.35864},
}