Table 10.

Results of an instrument detection task run on the RWC Instrument dataset.

Model	Accuracy	Macro-f1
Random Forest71${}^{\ast }$	.549	–
Partials72${}^{\ast }$	–	.634
CLMR26	.730	.717
VGGish	.821	.735
Cross-Dataset73${}^{\ast }$	–	.823
MFCCs	.913	.875
M2BERT, no pre-training	.930	.898
M2BERT	.954	.933
M3BERTSmall	.951	.912
M3BERTLarge	.966	.940

${}^{\ast }$ indicates that the model evaluates on different subsets of the dataset than our work and hence numbers are not directly comparable.

Highest values per metric are given in bold.