Tweaking the thinking instructions?

#104
by DaviKar - opened

Has anyone tried tweaking the thinking instructions for Qwen3.8?
We all know it takes years for Qwen3.8 to finish thinking, while low is fast there's quite a quality difference between low and xhigh (in many cases).
I was thinking if it's possible if there could be a good thinking instruction that can get close to xhigh, but just not overthink as much?

This finetune got released the other day!
https://huggingface.co/ukisai/Swift-Qwen3.8-27b

Yeah, but that's a finetune. I just mean editing the chat template.

Sign up or log in to comment