Can we get any Meta Data on which songs were in the Training Data?

[Cortexelus]

Can we get a list?

I'm curious about class imbalances, mis-labeled genres, ambiguous names that belong to multiple artists, artists with too few examples, artists with mixed live and studio recordings, etc, and how much these things are affecting output quality.

Discovery: With 7000+ labels this would help us narrow down which labels work better (sound more like the artist/genre, attend to the lyrics better, etc)

If we get a list I could do a quick meta-analysis.