LAVIS
LAVIS copied to clipboard
A reminder and question about the vicuna checkpoint
As a reminder, I find that the config of eachadea/vicuna-7b-1.1 and lmsys-vicuna-7b-v1.1 are different, i.e. they have different bos_token_id, eos_token_id, and pad_token_id, and only eachadea/vicuna-7b-1.1 can work well with instructBLIP.
By the way, I have a question about that: why these special token_ids are different but the model weight and other files are the same between eachadea/vicuna-7b-1.1 and lmsys-vicuna-7b-v1.1? In my opinion, if the token_id is changed, the model needs to be retrained. I would be appreciated if anybody could explain this :)
Are there any update to this issue?
same question! How to deal with it??