About performance of LLama (llama2, llama3) model

imoneoi / openchat

OpenChat: Advancing Open-source Language Models with Imperfect Data

https://openchat.team

Apache License 2.0

5.26k stars 398 forks source link

About performance of LLama (llama2, llama3) model #215

Open huazhenliu opened 6 months ago

huazhenliu commented 6 months ago

Thank you for your wonderful work!

Have you ever experimented with LLama2-7B as the model to do C-RLFT? How about the performance? Because OpenChat-3.5-0106 is based on Mistral, performance is really high, I have tried using LLama2-7B, the performance is not satisfied.

Another 2 questions: can chat model be used as the model to do C-RLFT? I think, some code needs to be done, e.g., chat template, etc. How about LLama3-8B-instuct, how to easy train, any performance data?

Thanks in advance.

imoneoi commented 6 months ago

Hi @huazhenliu We've tried Llama 2 13B, the performance is worse than Mistral 7B, so we've chosen Mistral 7B as the base model.

For your second question, it's OK. We can do C-RLFT on any model. You can edit the chat template here https://github.com/imoneoi/openchat/blob/master/ochat/config/__init__.py
We're actively working on a new version based on Llama-3-8B