RLHFlow / Qwen2.5-7B-DPO

Online-DPO-R1

• Blog : https://www.notion.so/Online-DPO-R1-1908b9a70e7b80c3bc83f4cf04b2f175
• Authors :
• Code : https://github.com/RLHFlow/Online-DPO-R1

Introduction

We release unofficial checkpoints for PPO, iterative DPO and rejection sampling (RAFT) trained from Qwen2.5-MATH-7B-base with rule-based RL, which are based on the success of Deepseek-R1-Zero and recent replications of PPO approach. Evaluated on five widely-adopted benchmarks AIME 2024 , MATH 500 , AMC , Minerva Math , OlympiadBench , our iterative DPO and RAFT model achieve significant enhancement compared to the base model and are comparable to the PPO approach. Our models are trained by using the prompt set from the MATH training set and Numina Math.

Moreover, we provide a detailed recipe to reproduce the model. Enjoy!

Model Releases

• [PPO model] ( https://huggingface.co/RLHFlow/Qwen2.5-7B-PPO-Zero )
• [Iterative DPO from SFT model] ( https://huggingface.co/RLHFlow/Qwen2.5-7B-DPO )
• [Iterative DPO from base model] ( https://huggingface.co/RLHFlow/Qwen2.5-7B-DPO-Zero )
• [Iterative DPO with Negative Log-Likelihood (NLL)] ( https://huggingface.co/RLHFlow/Qwen2.5-7B-DPO-NLL-Zero )
• [Raft] ( https://huggingface.co/RLHFlow/Qwen2.5-7B-RAFT-Zero )

Dataset

Training methods

• Iterative DPO: Following the RLHF Workflow framework ( https://arxiv.org/pdf/2405.07863 ), in each iteration, we sample multiple responses from the last trained policy, rank them via the ruled-based reward, and construct the preference pairs. Then, we optimize the policy by minimizing the DPO loss and enter the next iteration. Online iterative DPO can mitigate the issue of distribution shift and the limited coverage of offline data effectively
• RLHFlow/Qwen2.5-7B-DPO is trained based on the fine-tuned model Qwen2.5-Math-7B-Base + SFT Warm-up.

More detailed can be found in our blog !

Performance

Usage

Citation