The evaluation split of UniVerse (同谣).Training data lives in UniVerseSet.
UniVerseBench is a multilingual folk-music understanding benchmark for large audio–language models (LALMs). It asks models to listen, not to guess from language priors.
「诗言志，歌永言，声依永，律和声。」—《尚书·舜典》
Sister dataset (training)
universe-team/universeset
Two views of the same benchmark:
