Finetuning GTP-OSS 20b -- Does it need an actual 'Thinking' response?

Posted by searstream@reddit | LocalLLaMA | View on Reddit | 4 comments

Just seeing if anyone has ideas on if I can just set every "thinking" parameter to "Null" or if I really need to give it a reason why it should chose that answer. Thanks in advance.

Reply to Post

4 Comments

[-]

TimpaSurf@reddit

Yeah you will unlearn the "thinking" capability. But why finetune that model then? You could take e.g \`mistralai/Mistral-Small-3.2-24B-Instruct-2506\` instead.

[-]

searstream@reddit (OP)

Thanks, I am using both models. GPT just starts closer to my goal.

[-]

TimpaSurf@reddit

Have you found GPT-OSS-20B any better than Mistral24B?

[-]

searstream@reddit (OP)

Yes, in that it finds a lot less false positives. But missed more obvious ones. So they both miss the mark, but one goes over and the other under.