Provided tuning script maybe error
#59
by
efei - opened
for trl script, compute loss use all tokens exclude <pad>
for colab script, compute loss use all tokens exclude <pad> <image>
there are also <fake_image_token> and user turn should not be computed.
that's indeed correct! good catch @efei
@edbeeching can we change your trl gist?
Niels fixed a discrepancy earlier this week: https://github.com/huggingface/transformers/pull/30898#issuecomment-2124884284
efei changed discussion status to
closed