This is a process-supervised reward (PRM) from the project RLHFlow/RLHF-Reward-Modeling
The model is trained from meta-llama/Llama-3.1-8B-Instruct on RLHFlow/Deepseek-PRM-Data for 1 epochs. We use a global batch size of 32 and a learning rate of 2e-6, where we pack the samples and split them into chunks of 8192 token. See more training details at https://github.com/RLHFlow/RLHF-Reward-Modeling/tree/main/math-rm.
BoN evaluation result for Mistral generator:
| Model | Method | GSM8K | MATH |
|---|---|---|---|
| Mistral-7B | Pass@1 | 77.9 | 28.4 |
| Mistral-7B | Majority Voting@1024 | 84.2 | 36.8 |
| Mistral-7B | Mistral-ORM@1024 | 90.1 | 43.6 |
| Mistral-7B | Mistral-PRM@1024 | 92.4 | 46.3 |
Scaling the inference sampling to N=1024 for Deepseek generator:
| Model | Method | GSM8K | MATH |
|---|---|---|---|
| Deepseek-7B | Pass@1 | 83.9 | 38.4 |
| Deepseek-7B | Majority Voting@1024 | 89.7 | 57.4 |
| Deepseek-7B | Deepseek-ORM@1024 | 93.4 | 52.4 |
| Deepseek-7B | Deepseek-PRM@1024 | 93.0 | 58.1 |
| Deepseek-7B | Mistral-ORM@1024 (OOD) | 90.3 | 54.9 |
| Deepseek-7B | Mistral-PRM@1024 (OOD) | 91.9 | 56.9 |
Visualization
Usage
See https://github.com/RLHFlow/RLHF-Reward-Modeling/tree/main/math-rm for detailed examples.
Citation
The automatic annotation was proposed in the Math-shepherd paper:
@inproceedings{wang2024math,
title={Math-shepherd: Verify and reinforce llms step-by-step without human annotations},
author={Wang, Peiyi and Li, Lei and Shao, Zhihong and Xu, Runxin and Dai, Damai and Li, Yifei and Chen, Deli and Wu, Yu and Sui, Zhifang},
booktitle={Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
pages={9426--9439},
year={2024}
}
If you find the training recipe useful, please consider cite it as follows.
@misc{xiong2024implementation,
title={An implementation of generative prm},
author={Xiong, Wei and Zhang, Hanning and Jiang, Nan and Zhang, Tong},
year={2024}
}