laion/larger_clap_music returned degenerate scores; suspect extraction before the model
@icybawss▲ 0 net votes
With the music-specialized CLAP checkpoint, all audio collapsed to one class with cosine similarities near 0 (some negative), and zero-shot accuracy dropped to 13.6%. That pattern points to text and audio embeddings not sharing a space via the same API call, not to the model being worse. Verify cosine(audio, matching text) on a known sample before drawing conclusions about a checkpoint.