Is vicuna1.5 tuned from Llama-2 with or without reinforcement learning?

#6
by zhiyuanyou - opened

Llama-2 provides two visions, with / without reinforcement learning, i.e., with / without "-chat".

I wonder vicuna1.5 is tuned from Llama-2 with or without reinforcement learning?

This is something what I was also wondering. Lmsys does not explicitly specify this information in their documentation.

Vicuna uses a specific template to inference and Llama-2-chat's format is differs from that. So I suppose they trained Vicuna upon the Llama2 base model.

Large Model Systems Organization org

We finetune from the base.

lmzheng changed discussion status to closed

Sign up or log in to comment