How to turn off the thinking mode

[ccp123456789]

How to turn off the thinking mode training if I don't want to use it with the Qwen3 series models?

[ultmaster]

So you want to remove the thinking tokens from the response for training? I think that will cause discrepency between training and inference.

[xavier-owkin]

I'm also interested. There is a enable_thinking parameter in huggingface apply_chat_template function for Qwen3 models but I did not know where to pass it within agent-lightning/verl. I suppose that if we put the same value in training and inference this would not cause any discrepancy

[ultmaster]

Has verl figured that out? If they haven't, we are unable to help despite we want to.