Closed HackGiter closed 4 days ago
Currently, stop word of template of qwen is <|im_end|>. I think it should be <|endoftext|>, right? Normal behavior of pretraining data processing should be every example is seperated by <|endoftext|> instead of <|im_end|>
Do not specify the template argument during pretraining. It will use <|endoftext|> as the eos token
template
<|endoftext|>
Currently, stop word of template of qwen is <|im_end|>. I think it should be <|endoftext|>, right? Normal behavior of pretraining data processing should be every example is seperated by <|endoftext|> instead of <|im_end|>