What is the instruction format used for training?
If I like to add another fine tune on top of this, what instruction format am I supposed to use? I am slightly confused because I thought mistral 7B is using in general the one with [INST], but here (https://huggingface.co/cognitivecomputations/dolphin-2.8-mistral-7b-v02/blob/main/added_tokens.json) I see added tokens <|im_start|> and <|im_end|>
Is it:
<s>[INST] %input% [/INST]
%output%</s>
Or:
<|im_start|>system
%system%
<|im_end|>
<|im_start|>user
%input%
<|im_end|>
<|im_start|>assistant
%output%
<|im_end|>
When people fine-tune models, they always use the base model, which is not instruct-tuned. Mistral's fine-tunes of their models use their own instruct format, which uses the [INST] and [/INST] tokens. All of Eric's fine-tunes use ChatML for the instruct format.
If you're wanting to continue fine-tuning this model, then you would need to use ChatML. If you're wanting to fine-tune Mistral-7b-V0.2, the base model, then you can use any instruct format you want.