Step-DPO: Step-wise Preference Optimization for Long-chain Reasoning of LLMs
This repo contains the DeepSeekMath-RL-Step-DPO model. It is obtained by performing Step-DPO on DeepSeekMath-RL.
Step-DPO is a simple, effective, and data-efficient method for boosting the mathematical reasoning ability of LLMs. Notably, Step-DPO, when applied to Qwen2-72B-Instruct, achieves scores of 70.8% and 94.0% on the test sets of MATH and GSM8K without bells and wistles, respectively, surpassing a series of closed-source models, including GPT-4-1106, Claude-3-Opus, and Gemini-1.5-Pro.