Wrong Tokenizer?

#2
by anas-awadalla - opened

It seems like the tokenizer for this model is not correct as in I get very bad generation results if I use the tokenizer from OPT-30B model. In my case I am doing in-context learning evaluations and I get expected performance when I use the tokenizer from the 30B model. I am wondering if the tokenizer linked to this model is incorrect as I also see that the 30B model had a fix (https://huggingface.co/facebook/opt-30b/discussions/1) merged that I don't see for this model.

Hey @anas-awadalla ,

Thanks a lot for you message!
Could you provide me with a code snippet that shows why you think the tokenizer is incorrect?

The following code of the README seems to work well for me: https://huggingface.co/facebook/opt-1.3b#how-to-use

anas-awadalla changed discussion status to closed

Sign up or log in to comment