microsoft

/

Phi-3-mini-4k-instruct

like 635

Hi,

I noticed you changed the tokenizer in the last few days to ignore \n after <|user|>, <|assistant|>, <|system|> tokens #53. I feel that the model has trouble following instructions with this change, I would like to ask why the change in the template? Did you find that the model performs better with this change?

Thanks