Open source code for RL training in large language models.
A 7B Chinese reward model based on openChineseLlama.
A 7B English reward model based on Llama-7B.
SFT model for English.
Policy model for English after RLHF.
...
🌠 Introduction
Due to the challenges of reward design, environment interaction, and agent training, coupled with huge trial and error cost of large language models, there is a significant barrier for AI researchers to motivate the development of technical alignment and safe landing of LLMs. The stable training of RLHF has still been a puzzle.
In this technical report, we intend to help researchers to train their models stably with human feedback.
Contributions are summarized as follows:
We release competitive Chinese and English reward models, respectively, which have good cross-model generalization ability, alleviating the cost of relabeling human preference data;
We conduct in-depth analysis on the inner workings of PPO algorithm and propose the PPO-max algorithm to ensure stable model training;
We release the complete PPO-max codes to ensure that the LLMs in the current SFT stage can be better aligned with humans.
🔩 Requirements & Setup
This repository works on Python 3.8 and PyTorch 1.13.1.
We recommend using the conda virtual environment to run the code.
We can not directly release the full weight of the reward model because of protocol restrictions.
You can merge the diff weight with original Llama-7B to recover the reward model we used.
We upload the diff models, thanks to tatsu-lab, you can recover the reward model follow these steps:
bash
11) Download the weight diff into your local machine. The weight diff is located at:
2# For English:3TODO
4# For Chinese:5https://huggingface.co/Ablustrund/moss-rlhf-reward-model-7B-zh/tree/main
672) Merge the weight diff with the original Llama-7B:
8# For English:9# Reward model10python merge_weight_en.py recover --path_raw decapoda-research/llama-7b-hf --path_diff ./models/moss-rlhf-reward-model-7B-en/diff --path_tuned ./models/moss-rlhf-reward-model-7B-en/recover --model_type reward
11# SFT model12python merge_weight_en.py recover --path_raw decapoda-research/llama-7b-hf --path_diff ./models/moss-rlhf-sft-model-7B-en/diff --path_tuned ./models/moss-rlhf-sft-model-7B-en/recover --model_type sft
13# Policy model14TODO
15# For Chinese:16python merge_weight_zh.py recover --path_raw decapoda-research/llama-7b-hf --path_diff ./models/moss-rlhf-reward-model-7B-zh/diff --path_tuned ./models/moss-rlhf-reward-model-7B-zh/recover
Step 2: Select your own SFT model.
Because of some limitations, we can not release the Chinese SFT model (Currently).
You can use your own SFT model, or a strong base model instead of our SFT model.
Step 3: Start training
Run the command below.
# For Chinese:
# You need to use your own sft model currently.
bash run_zh.sh
# For English:
# We have loaded the sft model and reward model to huggingface.
bash run_en.sh
Citation
bibtex
1@article{zheng2023secrets,
2 title={Secrets of RLHF in Large Language Models Part I: PPO},
3 author={Rui Zheng and Shihan Dou and Songyang Gao and Wei Shen and Binghai Wang and Yan Liu and Senjie Jin and Qin Liu and Limao Xiong and Lu Chen and Zhiheng Xi and Yuhao Zhou and Nuo Xu and Wenbin Lai and Minghao Zhu and Rongxiang Weng and Wensen Cheng and Cheng Chang and Zhangyue Yin and Yuan Hua and Haoran Huang and Tianxiang Sun and Hang Yan and Tao Gui and Qi Zhang and Xipeng Qiu and Xuanjing Huang},
4 year={2023},
5 eprint={2307.04964},
6 archivePrefix={arXiv},
7 primaryClass={cs.CL}
8}
