# Light-R1: Curriculum SFT, DPO and RL for Long COT from Scratch and Beyond

**Authors:** Liang Wen, Yunke Cai, Fenrui Xiao, Xin He, Qi An, Zhenyu Duan, Yimin Du, Junchen Liu, Lifu Tang, Xiaowei Lv, Haosheng Zou, Yongchao Deng, Shousheng Jia, Xiangzheng Zhang  
**arXiv:** [2503.10460](https://arxiv.org/abs/2503.10460)  
**Date:** 2025-03-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2503.10460v4/

## Key points

- Light-R1 is an open-source suite for training long reasoning models in a reproducible and cost-effective manner
- Curriculum training with increasing data difficulty and multi-staged post-training techniques are key components of the methodology
- The Light-R1-32B model outperforms DeepSeek-R1-Distill-Qwen-32B in math reasoning
- Fine-tuning DeepSeek-R1-Distilled models with 3,000 challenging examples from the curriculum dataset leads to state-of-the-art 7B and 14B models
- The final model, Light-R1-14B-DS, achieves state-of-the-art performance in math with AIME24 & 25 scores surpassing many other models
- Light-R1 demonstrates strong cross-domain generalization capabilities
- Models, training data, and code are openly available at https://github.com/Qihoo360/Light-R1

## Summary

, , , , 


In this paper, we introduce Light-R1, an open-source suite designed for training long reasoning models in a reproducible and cost-effective manner. Our methodology involves curriculum training that gradually increases the difficulty of the data, coupled with multi-staged post-training techniques. The Light-R1-32B model demonstrates superior performance in math reasoning compared to DeepSeek-R1-Distill-Qwen-32B. Through experimental results, we show that our curriculum approach is most effective when diverse datasets are available for different training stages. Fine-tuning DeepSeek-R1-Distilled models with 3,000 challenging examples from our curriculum dataset has led to state-of-the-art 7B and 14B models. Additionally, the 32B model, Light-R1-32B-DS performs comparably to QwQ-32B and DeepSeek-R1. Furthermore, we extend our research by implementing GRPO on long reasoning models. Our final model, Light-R1-14B-DS achieves state-of-the-art performance among 14B models in math with AIME24 & 25 scores of 74.0 and 60.2 respectively, surpassing many 32B models and DeepSeek-R1-Distill-Llama-70B. Despite its focus on math training, Light-R1-14B-DS showcases strong cross-domain generalization capabilities. Overall, Light-R1 represents a significant advancement in making sophisticated reasoning models more accessible and implementable in real-world applications. Our models, training data, and code are openly available at https://github.com/Qihoo360/Light-R1 for further exploration and implementation purposes.