Ask about the usage of template

[coranholmes]

Thanks for the awesome work! I am not sure whether I have missed this.

I see from here, the train.py uses vicuna template.

FastChat/fastchat/train/train.py

Lines 87 to 88 in c3ad73a

	conv = get_conversation_template("vicuna")
	roles = {"human": conv.roles[0], "gpt": conv.roles[1]}

Do I need to change the template to "llama-2" if I would like to finetune llama2 model? I am a bit confused as if I set it to "llama-2", the training loss is always 0, but if I stick to "vicuna", the loss is okay.
I try to take a look at train_baichuan.py, it also uses "vicuna" instead of "baichuan..."

FastChat/fastchat/train/train_baichuan.py

Lines 80 to 82 in c3ad73a

	def apply_prompt_template(sources, systems=None):
	conv = get_conversation_template("vicuna")
	roles = {"human": conv.roles[0], "gpt": conv.roles[1]}

In that case, when should these templates be used?

[akujuou-sony]

I am facing the same issue with fine-tuning Mistral. Did anyone figure this out yet?

[akujuou-sony]

Figured it out not long after asking. For those that face this, you do need to modify the preprocess function in train.py. Specifically the # Mask targets. Only compute loss on the assistant outputs. sep = conv.sep + conv.roles[1] + ": " for conversation, target in zip(conversations, targets): and make sure it aligns with the template you use. I had to change it with ideas from #2423