stop the fine-tuned model to keep generating tokens

#45
by AlbelTec - opened

Hello,

I fine-tuned phi-1.5 but when trying to infer a question I'm getting the answer and the model keep generating tokens until it reach max_lenght tokens. As a newbie, I wonder how to prevent such behavior. any insights ?

kr,

Microsoft org

Are you setting the EOS token? What template did you use on your finetune data?

Actually, I used a subset of openorca dataset and my template is based on :
system_prompt:
question:
answer:

AlbelTec changed discussion status to closed

You ever find an answer to this question?

Sign up or log in to comment