quinnlue/audioset24k-mae-hardness

AudioSet-Opus MAE reconstruction hardness Per-clip MAE reconstruction loss over danjacobellis/audioset_opus_24kbps, computed with a frozen dasheng-base encoder + trained MAE decoder. Loss is the per-patch normalised reconstruction MSE on masked, in-span patches, averaged over 4 fixed-seed masking passes. Higher loss = harder example. Columns: id, path, loss, loss_all_patches, n_valid_patches, duration_s, labels, rank (1 = hardest).