Does anyone feel Qwen3 often fails to follow instructions accurately?

#18
by DOFOFFICIAL - opened

Does anyone feel Qwen3 often fails to follow instructions accurately?

Like, I asked it to write code to numerically solve some problems but it always tries to mathematically derive it. I asked to numerically solve since it does not exist an analytical solution.
But qwen3 just keeps reasoning and saying "Well,,," "But,..." "Wait,,,"

It's like,,,, benchmark oriented? :( like llama4?

i think qwq 32b is much better in instruction following

I feel it is really good at instructions as I have been testing on my rag chatbot project. It performs really well when i make the instruction more precise.
Previously I used Mistral-7B-Instruct-v0.3 which is like dogsh*t compared to qwen3 when it comes to answers.

"i think qwq 32b is much better in instruction following"

Yeah, same for me

@DOFOFFICIAL I have the same problems and QwQ seems to be much better in complex reasoning. Qwen3 have similar problems as QwQ-Preview. Please check my topic: https://huggingface.co/Qwen/Qwen3-235B-A22B/discussions/32

in either case, it is highly recommended to use both, since they have both pros and cons. for now, both of these models topple that of deepseek r1-distill-32b (esp. on huggingchat!), since the latter hallucinates too much.
https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B/discussions/45

Sign up or log in to comment