Prefix-RFT

This is the codebase of the paper: Blending Supervised and Reinforcement Fine-Tuning with Prefix Sampling. In this paper, we propose a hybrid approach for LLM post-training.

Structure

The code base in based on veRL. All prefix-rft related codes are in recipe/prefix_rft

Cite us

@article{huang2025blending,
    title={Blending Supervised and Reinforcement Fine-Tuning with Prefix Sampling},
    author={Huang, Zeyu and Cheng, Tianhao and Qiu, Zihan and Wang, Zili and Xu, Yinghui and Ponti, Edoardo M and Titov, Ivan},
    journal={arXiv preprint arXiv:2507.01679},
    year={2025}
}