News for hackers. Stuff that matters.

A comment on GPT-OSS Reinforcement Learning. Parent comment

Re:GPT-OSS Reinforcement Learning by BoorishBears on Saturday September 27, 2025 @04:35AM (#45394110)
Use constrained generation
[ Parent | on HN ]
Re:GPT-OSS Reinforcement Learning by artdigital on Saturday September 27, 2025 @06:11AM (#45394528)
Mind explaining?
[ Parent | on HN ]
Re:GPT-OSS Reinforcement Learning by 7thpower on Saturday September 27, 2025 @07:50AM (#45394960)
Curious as well
[ Parent | on HN ]
Re:GPT-OSS Reinforcement Learning by BoorishBears on Saturday September 27, 2025 @03:05PM (#45398538)
If you constrain the model to a JSON schema, most frivolous refusals go away.

And if you finetune on a few formatted examples the effect is even greater

[ Parent | on HN ]
Re:GPT-OSS Reinforcement Learning by corlinp on Saturday September 27, 2025 @12:11PM (#45397040)
Do you mean like structured outputs? Unfortunately here the model is guided to explicitly tell you when you violate the rules and why, it can confuse it's system rules with the game rules and say you're not allowed to ask a question about copyrighted material etc.
[ Parent | on HN ]