SA stands for Safety and alignment. We fine tuned DeepCoder-1.5B-Preview with STAR-1 for 250 steps to enhance safety alignment using unsloth SFT cookbook.
This model is fine-tuned with policy-grounded data to be safe and aligned with human values while coding. Specifically, it utilizes the STAR-1 dataset, which integrates diverse, deliberative reasoning examples evaluated rigorously by GPT-4o. This ensures the model maintains robust safety standards and minimizes biases, promoting responsible, secure, and effective coding practices without compromising its core reasoning capabilities.
1@misc{deepcoder2025,
2 title={DeepCoder: A Fully Open-Source 14B Coder at O3-mini Level},
3 author={Michael Luo, Sijun Tan, Roy Huang, Ameen Patel, Alpay Ariyak, Qingyang Wu, Xiaoxiang Shi, Rachel Xin, Colin Cai, Maurice Weber, Ce Zhang, Li Erran Li, Raluca Ada Popa, Ion Stoica},
4 howpublished={\url{https://pretty-radio-b75.notion.site/DeepCoder-A-Fully-Open-Source-14B-Coder-at-O3-mini-Level-1cf81902c14680b3bee5eb349a512a51}},
5 note={Notion Blog},
6 year={2025}
7}
This qwen2 model was trained 2x faster with
Unsloth and Huggingface's TRL library.
