VA-Judger Dataset
This public dataset repository bundles the VA-Judger benchmark and its decontaminated human-preference training set.
Contents
VA-Judger-Dataset/
├── VA-Judger-Bench/
│ ├── easy/ # 400 cases
│ ├── indomain/ # 250 cases
│ └── outdomain/ # 500 cases
└── VA-Judger-Train/
├── train/ # preference pairs and videos
The benchmark contains 1,150 paired audio/video cases. Every benchmark record includes the human preference, the selected A-F reason codes, and any custom free-text reason.
The training set contains human preference labels, A–F reason codes, prompts, and two videos per pair. Exact case-ID, prompt-text, and SHA-256 video overlaps with all three benchmark splits were removed.
See the README inside each top-level directory for schemas, label definitions, statistics, curation details, and usage examples.
Human preference reason labels
The human-selected reason labels are multi-select:
- (A) Better match to the prompt (characters, content, or speech)
- (B) Better audio-visual consistency (e.g., lip sync, sound matches the image)
- (C) Higher audio quality
- (D) Higher video quality
- (E) More complete and coherent content
Licensing and responsible release
No redistribution license is asserted by this bundle. Dataset owners must verify video, prompt, likeness, voice, trademark, and copyright redistribution rights and add the intended license before relying on this repository for a formal public release.
- Downloads last month
- 156