An SFT training corpus that powers the AnyAudio-Judge evaluator. Each sample contains:
An audio clip (referenced relatively under audios/).
A multi-turn chat (messages) where the user enumerates a list of decomposed binary rubric items and the assistant answers them in JSON, with per-item evidence (Chain-of-Thought rationale).
A coarse label ("yes" if the caption originally matched the audio, "no" otherwise) and a tag describing how the caption was… See the full description on the dataset page:
https://huggingface.co/datasets/cucl2/AnyAudio-Judge-Corpus.
