Xiaoyun Wu

Do you have the numbers that shows method 2 works and better? @prithivMLmods

method 2 has vllm in it, but disabled, method 1 does on have vllm in it. am I missing something? So method 2 is the one to use? if enable grpo?

which one is vllm based? How can one tell? can you mention it in article?
Aslo, are you happen to aware work on using grpo to improve MMLU (or some task inside it) with models like qwen 2.5 7b or even smaller? @prithivMLmods thanks.

why there are 2 methods?