Inspired by the success of
DeepSeek-R1 in training models using RL, many open-source projects have made significant progress in exploring the effectiveness of RL training. These projects not only validate RL as an effective training method but also showcase its potential across various application scenarios. However, such efforts have so far only replicated this success
within limited domains and
with smaller-parameter LLMs, without fully extending to larger-scale, more complex models and a broader range of tasks.
Med-R1 is dedicated to translating the success of RL in the training of LLMs within general domains to the medical field. To ensure that the model possesses a comprehensive reserve of medical knowledge, we have adopted a large-parameter base model. Specifically, this model integrates vast amounts of multi-source heterogeneous data, including medical literature, clinical guidelines, and electronic health records during the pre-training phase. Through fine-tuning, the model is refined to accurately understand and generate specialized medical content with nuanced reasoning.
1@Misc{med-r1,
2 title = {Med-R1: Encourage Medical LLM to engage in deep thinking similar to DeepSeek-R1},
3 author = {Rongsheng Wang},
4 howpublished = {\url{https://github.com/WangRongsheng/Med-R1}},
5 year = {2025}
6}
