Layer-wise downstream results for 17 music foundation models. Each cell shows
the best layer's score, its index, and the full layer profile as a sparkline.
Click any column header to sort; family headers (▸) expand companion metrics.
Rank = mean of the model’s per-task ranks (1 = best;
p ranked on fewer tasks)
· Params = evaluated encoder (audio tower for two-tower; decoder for AR)
· + beyond the paper’s 12 models
· † metric caveat (hover) · — not evaluated
· dot = best layer