Uploaded my Qwen3.6 27B based fine tune, after two years of experience fine tuning models

Posted by de4dee@reddit | LocalLLaMA | View on Reddit | 14 comments

Been doing fine tuning for more than 2 years, using different tools but mostly Unsloth.

I think my dataset expansion tricks worked better this time, it reached 75% human alignment compared to 73% for previous Qwen 3.5 fine tune.

Btw I benchmark against my own evals, because not many people doing this kind of work.

If you try it, I would love some feedback. Thanks <3