What it is
2,800 audio-video dialogues with tiered rubrics for benchmarking native audio-visual chat, plus a shared reward recipe for RL training. 2,800 audio-video dialogues across 5 abilities, 17 subcategories, and 22 domains, with tiered rubrics for single- and multi-turn evaluation plus a shared reward implementation for RL training.
Key results
- 2,800 dialogues: 2,550 single-turn + 250 multi-turn
- 5 abilities, 17 subcategories, 22 scenario domains
- 3,490 1080p MP4s, ~28 GB, CC-BY-NC-ND-4.0
Useful for benchmarking and reward-model training for models that answer directly from synchronized video and audio.