Missing Chat Template
This meta-llama/Meta-Llama-3-8B model currently lacks a specified chat template in its repository. This absence could potentially lead to performance degradation due to incorrect chat formatting when the model is utilized.
Related Model
It's noted that the similar model meta-llama/Meta-Llama-3-8B-instruct has a defined chat template.
Suggestion
Could the chat template from meta-llama/Meta-Llama-3-8B-instruct be appropriate to apply to meta-llama/Meta-Llama-3-8B? If this is not suitable, could we add a specific chat template tailored for the meta-llama/Meta-Llama-3-8B model to prevent any potential formatting issues?
This is a base model, hence it's not conversational and does not have chat capabilities. The Instruct model you mention is conversational (and hence why it does have a template). The template is usually added during the fine-tuning process (which is when this base model became the instruct one).
Thanks for the clarification @osanseviero ! This makes sense.
My goal is to fine-tune Llama-3-8B to do multi-label classification on the MimicIII medical codes dataset and I am a bit lost on what prompt template to use. I am a bit paranoid about getting all the special tokens right. Would be awesome if you could point me into the right direction.
I realize this may be veering off the original discussion topic, so I'm happy to start a new thread if that would be more appropriate. Thanks for any assistance you can offer!