Ablustrund commited on
Commit
287f7f9
β€’
1 Parent(s): 20205a2

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +123 -0
README.md CHANGED
@@ -1,3 +1,126 @@
1
  ---
2
  license: agpl-3.0
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: agpl-3.0
3
+ language:
4
+ - zh
5
+ tags:
6
+ - llm
7
+ - moss
8
+ - rlhf
9
+ - policy model
10
  ---
11
+
12
+ # MOSS-RLHF
13
+
14
+ ### *MOSS-RLHF & "Secrets of RLHF in Large Language Models Part I: PPO" <br>πŸ‘‰ <a href="https://arxiv.org/abs/2307.04964" target="_blank">[Technical report]</a> <a href="https://openlmlab.github.io/MOSS-RLHF/" target="_blank">[Home page]*
15
+
16
+
17
+ ## 🌟 News
18
+ ### πŸ‘‰ Wed, 12. July 2023. We have released Chinese reward model based OpenChineseLlama-7B!
19
+ [moss-rlhf-reward-model-7B-zh](https://huggingface.co/Ablustrund/moss-rlhf-reward-model-7B-zh/tree/main)
20
+ <br>
21
+
22
+ ### πŸ‘‰ Thu, 13. July 2023. We have released English reward model and SFT model based Llama-7B!
23
+ [moss-rlhf-reward-model-7B-en](https://huggingface.co/fnlp/moss-rlhf-reward-model-7B-en)
24
+
25
+ [moss-rlhf-sft-model-7B-en](https://huggingface.co/fnlp/moss-rlhf-sft-model-7B-en)
26
+ <br>
27
+
28
+ ## 🧾 Open-source List
29
+ - [x] Open source code for RL training in large language models.
30
+ - [x] A 7B Chinese reward model based on openChineseLlama.
31
+ - [x] A 7B English reward model based on Llama-7B.
32
+ - [x] SFT model for English.
33
+ - [ ] Policy model for English after RLHF.
34
+ - ...
35
+
36
+ ## 🌠 Introduction
37
+
38
+ 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.
39
+ In this technical report, we intend to help researchers to train their models stably with human feedback.
40
+
41
+ Contributions are summarized as follows:
42
+ 1) We release competitive Chinese and English reward models, respectively, which have good cross-model generalization ability, alleviating the cost of relabeling human preference data;
43
+ 2) We conduct in-depth analysis on the inner workings of PPO algorithm and propose the PPO-max algorithm to ensure stable model training;
44
+ 3) We release the complete PPO-max codes to ensure that the LLMs in the current SFT stage can be better aligned with humans.
45
+
46
+
47
+ ## πŸ”© Requirements & Setup
48
+
49
+ This repository works on Python 3.8 and PyTorch 1.13.1.
50
+
51
+ We recommend using the **conda** virtual environment to run the code.
52
+
53
+ #### Step 1: Create a new Python virtual environment
54
+ ```bash
55
+ conda update conda -n base -c defaults
56
+ conda create -n rlhf python=3.8
57
+ conda activate rlhf
58
+ ```
59
+ #### Step 2: Install PyTorch and TensorBoard
60
+ ```bash
61
+ conda install pytorch==1.13.1 pytorch-cuda=11.7 tensorboard -c pytorch -c nvidia
62
+ ```
63
+
64
+ #### Step 3: Install the remaining dependencies
65
+ ```bash
66
+ conda install datasets accelerate safetensors chardet cchardet -c huggingface -c conda-forge
67
+ pip3 install transformers sentencepiece einops triton==1.0.0 rouge jionlp==1.4.14 nltk sacrebleu cpm_kernels
68
+
69
+ apt install libaio-dev
70
+ DS_BUILD_OPS=1 pip install deepspeed
71
+ ```
72
+
73
+ ## ✨ Start training your own model!
74
+ Run code in a few steps.
75
+
76
+ ### Step 1: Recover Reward model weights
77
+ We can not directly release the full weight of the reward model because of protocol restrictions.
78
+ You can merge the diff weight with original Llama-7B to recover the reward model we used.
79
+
80
+ We upload the diff models, thanks to tatsu-lab, you can recover the reward model follow these steps:
81
+ ```bash
82
+ 1) Download the weight diff into your local machine. The weight diff is located at:
83
+ # For English:
84
+ TODO
85
+ # For Chinese:
86
+ https://huggingface.co/Ablustrund/moss-rlhf-reward-model-7B-zh/tree/main
87
+
88
+ 2) Merge the weight diff with the original Llama-7B:
89
+ # For English:
90
+ # Reward model
91
+ python 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
92
+ # SFT model
93
+ python 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
94
+ # Policy model
95
+ TODO
96
+ # For Chinese:
97
+ python 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
98
+ ```
99
+ ### Step 2: Select your own SFT model.
100
+ Because of some limitations, we can not release the **Chinese** SFT model (Currently).
101
+ You can use your own SFT model, or a strong base model instead of our SFT model.
102
+
103
+ ### Step 3: Start training
104
+ Run the command below.
105
+ ```
106
+ # For Chinese:
107
+ # You need to use your own sft model currently.
108
+ bash run_zh.sh
109
+
110
+ # For English:
111
+ # We have loaded the sft model and reward model to huggingface.
112
+ bash run_en.sh
113
+ ```
114
+
115
+ ## Citation
116
+
117
+ ```bibtex
118
+ @article{zheng2023secrets,
119
+ title={Secrets of RLHF in Large Language Models Part I: PPO},
120
+ 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},
121
+ year={2023},
122
+ eprint={2307.04964},
123
+ archivePrefix={arXiv},
124
+ primaryClass={cs.CL}
125
+ }
126
+ ```