Qwen
/

Is it possible to decrease vocabulary size?

#50
by omers66 - opened

Is there a way to access the tokenizer tokens distribution? I would like to decrease vocabulary size to speed up. Would be ideal if I could keep most request tokens.
Thanks

You need to modify the embedding layer and the language modeling head as well.

Yes, but what about the histogram of tokens? I would like to remove the un-common ones...
Thanks

Sign up or log in to comment