What changes should I make to apply the SPIN method on Llama2?

[Labmem009]

I want to apply SPIN method on llama2 with alpaca-like finetuning datasets. What changes should I make to apply the SPIN method?
Thanks a lot!

[TanZhendong]

Hello. I've tried to apply SPIN on llama-2 with the ultrachat200k datasets. It seems that the tokenizer for llama2 and zephyr-7b are different (as shown in this document). As a result, the code in spin/run_spin.py likely needs to be modified (which is primarily focused on the apply_chat_template function). By the way, for alpaca-like datasets, the structure of the dataset is different from ultrachat200k (the key is "instruction" and "output" rather than "prompt" and "messages"). I believe just changing the code in spin/reformat.py to ensure the structure of the dataset is the same should be ok.