"bos_token": "<s>" vs. "<|endoftext|>"
#20
by
tanliboy
- opened
In the chat template of this model, it uses <|endoftext|>
as the bos_token instead of <s>
in other phi3 models.
As the text is "end of text", it is a little confusing. Is it intentional due to different token dataset? Do we expect a downgrade of performance if I replace the bos_token with <s>
?
Thank you for your feedback and interest in the Phi-3 model. We highly recommend users to follow with the suggested format in the model card.
nguyenbh
changed discussion status to
closed