Support thinking model

[aongwachi1]

Hi contributor, thanks for sharing your work.

I've read through the readme, it say "please select non-thinking model"
So it's mean the training script still not support thinking model like "qwen3-4b-thinking-2507" right?

[1GaryLi]

Sorry for missing the issue earlier and thanks for your interest in our work!

Yes, the current training scripts and synthetic SFT data are mainly for non-thinking models. Some of our synthetic trajectories don’t strictly follow the standard thinking-format for each turn (...[response]), so we fine-tune non-thinking models to avoid breaking the model’s original output format. Our experiments with Qwen3-8B were also done in the non-thinking setting. However, you could also try training a thinking model like qwen3-4b-thinking-2507 on our SFT data and it may still bring some performance gains.

Please let me know if you have any other questions!