Why does the CSV extracted according to the steps show all the scores as nan?

[SsmallSong]

file name, index, score
cos, 0, nan
cos, 1, nan
cos, 2, nan
cos, 3, nan
cos, 4, nan
cos, 5, nan
cos, 6, nan
cos, 7, nan
cos, 8, nan
cos, 9, nan
cos, 10, nan
cos, 11, nan
cos, 12, nan
cos, 13, nan
cos, 14, nan
cos, 15, nan
cos, 16, nan
cos, 17, nan
cos, 18, nan
cos, 19, nan
cos, 20, nan

[xiamengzhou]

I think it's largely because you were not getting the correct gradients from the models, could you check if the raw gradients before compression are correct?