Misleading chat template
Hey, first of all thank you very much for all your work and open-sourcing it. I've learned a lot from reading your technical report.
I noticed that the chat_template.jinja file includes templating with <|im_start|> and <|im_end|> tokens. My understanding is that the base models haven't been trained on this format so I would suggest either not having any template or a simple Question: <user message>\nAnswer: template instead. Same for the 32B base model of course.
Hi @ivoschaper , thanks so much for your kind words and feedback!
That's correct, the the base models should generally not have chat templates. This is a result of converting the checkpoints to Hugging Face format, which carried over a chat template from the tokenizer used for conversion. For now, I'm going to remove these files from both base models as to not cause confusion. Thanks for pointing this out!
Hi @baileyk , I'm happy to hear that my feedback was useful. Thank you for taking care of removing the files! :)