Missing ChatML special tokens
Hey, thanks for this model, the performance is great!
I noticed that the tokenizer config files are missing the ChatML special tokens like im_end - is this intentional?
Hi @brlambert ,
If you want to use chatml, you can just change the related tokenizer files and it should work fine.
Hi @brlambert ,
If you want to use chatml, you can just change the related tokenizer files and it should work fine.
Nice, I was hoping you'd say that! Thanks!
So I modified the files but I'm getting this error because of the addition of the <|im_start|>
and <|im_end|>
tokensValueError: Parameter model.embed_tokens.q_weight has shape (32000, 512), but expected (32002, 512)
Have you changed all tokenizer related files to match chatml?
Yes I believe so, but adding these new special tokens seems to be imcompatible with the model. I tried to fix this by editing the problematic layer shape with model.resize_token_embeddings(32002)
and I am able to generate outputs with this new model but it's all gibberish