Question about training set of CLAP

[sjhan91]

Hello, I wonder which dataset you used to train CLAP (especially for music).

The reason I ask is audio embeddings from audio syntheized from MIDI is not closely aligned with text embeddings (MusicCaps, AudioStock, LP-MusicCaps, AudioSet). (when I draw samples in t-SNE space)

Also, GTZAN embeddings show similar situation.

Or, can you let me know the example of captions?

Regards.