Upload README_hf.md with huggingface_hub
Browse files- README_hf.md +215 -0
README_hf.md
ADDED
@@ -0,0 +1,215 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
<div align="center">
|
2 |
+
|
3 |
+
# Open Reasoner Zero
|
4 |
+
|
5 |
+
<img src="figure/logo.jpg" width="300"/>
|
6 |
+
|
7 |
+
<div>
|
8 |
+
|
9 |
+
An Open Source Approach to Scaling Up Reinforcement Learning on the Base Model
|
10 |
+
</div>
|
11 |
+
</div>
|
12 |
+
|
13 |
+
<div align="center" style="line-height: 1;">
|
14 |
+
<a href="https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero" style="margin: 2px;"><img alt="Code" src="https://img.shields.io/badge/Open%20Reasoner%20Zero-000000?style=for-the-badge&logo=github&logoColor=000&logoColor=white" style="display: inline-block; vertical-align: middle;"/></a>
|
15 |
+
|
16 |
+
<a href="https://huggingface.co/Open-Reasoner-Zero" target="_blank"><img alt="Hugging Face"
|
17 |
+
src="https://img.shields.io/badge/HuggingFace-fcd022?style=for-the-badge&logo=huggingface&logoColor=000&labelColor"/></a>
|
18 |
+
|
19 |
+
<a href="https://yasminezhang.notion.site/Open-Reasoner-Zero-19e12cf72d418007b9cdebf44b0e7903" target="_blank">
|
20 |
+
<img alt="Notion Page"
|
21 |
+
src="https://img.shields.io/badge/Notion-%23000000.svg?style=for-the-badge&logo=notion&logoColor=white"/></a>
|
22 |
+
|
23 |
+
<br>
|
24 |
+
<a href="https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/blob/main/ORZ_paper.pdf"><b>Paper PDF Link [WIP]</b>ποΈ</a>
|
25 |
+
</div>
|
26 |
+
|
27 |
+
<div>
|
28 |
+
<br>
|
29 |
+
|
30 |
+
</div>
|
31 |
+
|
32 |
+
## Overview π
|
33 |
+
We introduce **Open-Reasoner-Zero**, the first open source implementation of large-scale reasoning-oriented RL training focusing on scalability, simplicity and accessibility.
|
34 |
+
|
35 |
+
To enable broader participation in this pivotal moment we witnessed and accelerate research towards artificial general intelligence (AGI),
|
36 |
+
we release our source code, parameter settings, training data, and model weights.
|
37 |
+
Please refer to our [paper](https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/blob/main/ORZ_paper.pdf) for more insights across various model sizes.
|
38 |
+
|
39 |
+
**Let the Reasoner-Zero tide rise!**
|
40 |
+
|
41 |
+
|
42 |
+
## Main Results π
|
43 |
+
|
44 |
+

|
45 |
+
|
46 |
+
*Figure 1 | Evaluation performance of Open-Reasoner-Zero-\{7B, 32B\}. Evaluation performance of Open-Reasoner-Zero-\{7B, 32B\} on benchmarks (averaged on 16 responses) during training. Using the same base model as DeepSeek-R1-Zero-Qwen-32B, Open-Reasoner-Zero-32B achieves superior performance on AIME2024, MATH500, and GPQA Diamond benchmark-requiring only a tenth of the training steps.*
|
47 |
+
|
48 |
+

|
49 |
+
*Figure 2 | Train-time Scale up on Train Reward and Response Length of Open-Reasoner-Zero (ORZ) - \{0.5B, 1.5B, 7B, 32B\}. Train Reward and Response Length increase steadily, demonstrating consistent scalability across model sizes. Interestingly, the ORZ-32B Response Length exhibits fluctuations without negatively impacting training stability, highlighting the robustness of our minimalist recipe.*
|
50 |
+
|
51 |
+
## Releases π¦
|
52 |
+
|
53 |
+
<strong>[2025/03/31]</strong>
|
54 |
+
We announce a major milestone for `Open-Reasoner-Zero`:
|
55 |
+
|
56 |
+
- π [Updated Paper](https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/blob/main/ORZ_paper.pdf) with new results.
|
57 |
+
- π [Easy-to-use Training Scripts](https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/tree/main/playground):
|
58 |
+
- [ORZ-1.5B training scripts](https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/blob/main/playground/orz_1p5b_ppo.py) and [ORZ-0.5B training scripts](https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/blob/main/playground/orz_0p5b_ppo.py) (main results in Figure 2).
|
59 |
+
- [Minimal resource training scripts](https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/blob/main/playground/orz_0p5b_ppo_1gpu.py): ORZ-0.5B can be run on a single A800/H800 gpu!
|
60 |
+
- π€© [Updated Curated Datasets](https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/tree/main/data):
|
61 |
+
- 129k data in total:
|
62 |
+
- [original 57k data](https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/blob/main/data/orz_math_57k_collected.json).
|
63 |
+
- [extended 72k data](https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/blob/main/data/orz_math_72k_collection_extended.json).
|
64 |
+
- [13k hard data](https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/blob/main/data/orz_math_13k_collection_hard.json) mined from the above 129k data.
|
65 |
+
- used in the "annealing" stage of ORZ-32B training: **AIME2024 from ~41% to ~48%**!
|
66 |
+
- π€ More HF Models:
|
67 |
+
- Updated HF Models: [`Open-Reasoner-Zero-7B`](https://huggingface.co/Open-Reasoner-Zero/Open-Reasoner-Zero-7B) and [`Open-Reasoner-Zero-32B`](https://huggingface.co/Open-Reasoner-Zero/Open-Reasoner-Zero-32B).
|
68 |
+
- Released HF Models: [`Open-Reasoner-Zero-1.5B`](https://huggingface.co/Open-Reasoner-Zero/Open-Reasoner-Zero-1.5B) and [`Open-Reasoner-Zero-0.5B`](https://huggingface.co/Open-Reasoner-Zero/Open-Reasoner-Zero-0.5B).
|
69 |
+
- π Full Suite of Critic Models for in-depth research: `Open-Reasoner-Zero-Critic-`{[0.5B](https://huggingface.co/Open-Reasoner-Zero/Open-Reasoner-Zero-Critic-0.5B), [1.5B](https://huggingface.co/Open-Reasoner-Zero/Open-Reasoner-Zero-Critic-1.5B), [7B](https://huggingface.co/Open-Reasoner-Zero/Open-Reasoner-Zero-Critic-7B), [32B](https://huggingface.co/Open-Reasoner-Zero/Open-Reasoner-Zero-Critic-32B)}.
|
70 |
+
|
71 |
+
<strong>[2025/02/18]</strong>
|
72 |
+
We release `Open-Reasoner-Zero`.
|
73 |
+
|
74 |
+
As part of this release, we open-source:
|
75 |
+
- π [Paper](https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/blob/main/ORZ_paper.pdf) on our comprehensive analysis and insights in Reasoner-Zero training
|
76 |
+
- π€ HF Model [`Open-Reasoner-Zero-7B`](https://huggingface.co/Open-Reasoner-Zero/Open-Reasoner-Zero-7B) and [`Open-Reasoner-Zero-32B`](https://huggingface.co/Open-Reasoner-Zero/Open-Reasoner-Zero-32B)
|
77 |
+
- π [`Our curated 57k training data`](https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/tree/main/data)
|
78 |
+
- π [Training Scripts](https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/tree/main/playground) to enjoy your own Reasoner-Zero journey!
|
79 |
+
|
80 |
+
## Key Features in Codebase π
|
81 |
+
|
82 |
+
- Adopt single controller trainer design, flexible and researcher-friendly.
|
83 |
+
- Colocate training and generation in the same GPUs to maximize GPU utilization.
|
84 |
+
|
85 |
+
## Getting Started π
|
86 |
+
### Data
|
87 |
+
|
88 |
+
We release all of curated high-quality training data in the [`data`](https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/tree/main/data) folder:
|
89 |
+
* curated 129k data:
|
90 |
+
* [original 57k](https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/blob/main/data/orz_math_57k_collected.json), collected from various sources, including AIME (up to 2023), MATH, Numina-Math collection and Tulu3 MATH.
|
91 |
+
* [extended 72k](https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/blob/main/data/orz_math_72k_collection_extended.json), mainly cleaned from OpenR1-Math-220k.
|
92 |
+
* [hard 13k](https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/blob/main/data/orz_math_13k_collection_hard.json), mined from the first stage of ORZ-32B training.
|
93 |
+
|
94 |
+
The details for how to collect data are described in our [paper](https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/blob/main/ORZ_paper.pdf).
|
95 |
+
|
96 |
+
### Installation & Training Scripts
|
97 |
+
We release our [Dockerfile](https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/blob/main/docker/Dockerfile) in [docker](https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero/tree/main/docker) folder to facilitate the reproducibility of our training.
|
98 |
+
|
99 |
+
To install the package, run:
|
100 |
+
```bash
|
101 |
+
pip install -e .
|
102 |
+
```
|
103 |
+
|
104 |
+
#### Start ORZ-32B PPO Training
|
105 |
+
Here are the starting commands in 16 nodes.
|
106 |
+
|
107 |
+
First on master node, run:
|
108 |
+
```bash
|
109 |
+
ray start --head
|
110 |
+
# you will see logging like:
|
111 |
+
# Next steps
|
112 |
+
# To add another node to this Ray cluster, run
|
113 |
+
# ray start --address='<master-node-ip>:<master-node-port>'
|
114 |
+
```
|
115 |
+
|
116 |
+
then on all other nodes, run:
|
117 |
+
```bash
|
118 |
+
ray start --address='<master-node-ip>:<master-node-port>' # <master-node-ip> and <master-node-port> are from above loggings!
|
119 |
+
```
|
120 |
+
|
121 |
+
finally on master node, just run:
|
122 |
+
```bash
|
123 |
+
python -m playground.orz_32b_ppo
|
124 |
+
```
|
125 |
+
Your training log will be shown in the master node terminal.
|
126 |
+
|
127 |
+
------
|
128 |
+
|
129 |
+
#### Start ORZ-0.5B PPO Training
|
130 |
+
You can start the ORZ-0.5B PPO training in single A800/H800 node:
|
131 |
+
```bash
|
132 |
+
python -m playground.orz_0p5b_ppo
|
133 |
+
```
|
134 |
+
|
135 |
+
You can even run in **a single A800/H800 gpu**:
|
136 |
+
```bash
|
137 |
+
python -m playground.orz_0p5b_ppo_1gpu
|
138 |
+
```
|
139 |
+
|
140 |
+
note: since we are not in multi-node setting, no `ray start` like logics are needed.
|
141 |
+
|
142 |
+
------
|
143 |
+
|
144 |
+
#### Start ORZ-7B PPO Training
|
145 |
+
|
146 |
+
Multi-node Training on 4 nodes:
|
147 |
+
```bash
|
148 |
+
# set up for multi-node training
|
149 |
+
ray start --head # on master node
|
150 |
+
ray start --address='<master-node-ip>:<master-node-port>' # then on other nodes
|
151 |
+
|
152 |
+
# then on master node, run:
|
153 |
+
python -m playground.orz_7b_ppo
|
154 |
+
```
|
155 |
+
|
156 |
+
Your training log will be shown in the master node terminal.
|
157 |
+
|
158 |
+
-----
|
159 |
+
|
160 |
+
#### Start ORZ-1.5B PPO Training
|
161 |
+
|
162 |
+
Multi-node Training on 2 nodes:
|
163 |
+
```bash
|
164 |
+
# set up for multi-node training
|
165 |
+
ray start --head # on master node
|
166 |
+
ray start --address='<master-node-ip>:<master-node-port>' # then on other nodes
|
167 |
+
# then on master node, run:
|
168 |
+
python -m playground.orz_1p5b_ppo
|
169 |
+
```
|
170 |
+
|
171 |
+
----
|
172 |
+
|
173 |
+
#### Debug Settings
|
174 |
+
In the code, we leave an environment variable `DEBUG_MODE` to run in debug setting for researcher to iterate. (Thought for now, we recommend using `python -m playground.orz_0p5b_ppo_1gpu` for debugging.)
|
175 |
+
|
176 |
+
The debug running command examples:
|
177 |
+
```bash
|
178 |
+
# NOTE: just for debug, not final setting!
|
179 |
+
|
180 |
+
## Debug command in a single GPU with `EleutherAI/pythia-14m`
|
181 |
+
DEBUG_MODE=True python -m playground.orz_14m_ppo_mini
|
182 |
+
## Debug command in a single node (8 GPUs) with `Qwen/Qwen2.5-7B`
|
183 |
+
DEBUG_MODE=True python -m playground.orz_7b_ppo
|
184 |
+
```
|
185 |
+
|
186 |
+
## Acknowledgements π
|
187 |
+
|
188 |
+
- This work was supported by computing resources and valuable feedback provided by [StepFun](https://www.stepfun.com/) and Tsinghua University.
|
189 |
+
- Our training framework is built on [OpenRLHF](https://github.com/OpenRLHF/OpenRLHF), [vllm](https://github.com/vllm-project/vllm), [DeepSpeed](https://github.com/deepspeedai/DeepSpeed) and [ray](https://github.com/ray-project/ray).
|
190 |
+
- Our model is based on [Qwen2.5 Series](https://qwenlm.github.io/blog/qwen2.5-llm/) of **base models**, including [Qwen2.5-0.5B](https://huggingface.co/Qwen/Qwen2.5-0.5B), [Qwen2.5-1.5B](https://huggingface.co/Qwen/Qwen2.5-1.5B), [Qwen2.5-7B](https://huggingface.co/Qwen/Qwen2.5-7B) and [Qwen2.5-32B](https://huggingface.co/Qwen/Qwen2.5-32B).
|
191 |
+
- We thank [Project Numina](https://projectnumina.ai/), [Tulu3](https://allenai.org/blog/tulu-3-technical) and [OpenR1-Math-220k](https://huggingface.co/datasets/open-r1/OpenR1-Math-220k) for their collected open sourced data.
|
192 |
+
|
193 |
+
## Advertisement Time π£
|
194 |
+
|
195 |
+
We are hiring talented researchers and engineers to join our team. If you are interested in our project and would like to contribute to the reasoner scale-up all the way to AGI, please feel free to reach out to us at [email protected]
|
196 |
+
|
197 |
+
|
198 |
+
[](https://star-history.com/#Open-Reasoner-Zero/Open-Reasoner-Zero&Timeline)
|
199 |
+
|
200 |
+
## Community Discussions πΊ
|
201 |
+
|
202 |
+
We have several wechat groups to help discussions and sharing, you can scan the QR code below to join the latest group.
|
203 |
+
|
204 |
+
<img src="figure/WeChatGroup.png" width="300" style="display: block; margin: 0 auto;"/>
|
205 |
+
|
206 |
+
## Citation
|
207 |
+
|
208 |
+
```bibtex
|
209 |
+
@misc{OpenReasonerZero2025,
|
210 |
+
title={Open-Reasoner-Zero: An Open Source Approach to Scaling Reinforcement Learning on the Base Model},
|
211 |
+
author={Jingcheng Hu and Yinmin Zhang and Qi Han and Daxin Jiang and Xiangyu Zhang, Heung-Yeung Shum},
|
212 |
+
year={2025},
|
213 |
+
howpublished={\url{https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero}},
|
214 |
+
}
|
215 |
+
```
|