which chat template should we use?

#6
by wyxwangmed - opened

Hello, we have largely conducted training without explicit use of a chat template so I think that the above template that @ehartford linked is the best option. We will run a few tests to verify this and update the model card/tokenizer_config accordingly soon.

Hi, we have conducted an experiment with two different chat templates on MT-Bench. The two were the Llama-2 chat template (essentially the MoMo one linked above) as well as the Qwen chat template from here: https://huggingface.co/Qwen/Qwen-7B-Chat/blob/main/qwen_generation_utils.py#L130. In both cases we found fairly close scores so we feel that either is OK, though Llama-2 performed a bit better.

For reference our average scores on MT-Bench across 2 turns are in the region of ~7.75

I tried both llama2 prompt and qwen72b prompt, nothing works. Always get garbage results.

Sign up or log in to comment