Model Overview
Model Features
Model Capabilities
Use Cases
đ Open Reasoner Zero
An open source approach to scaling up reinforcement learning on the base model, focusing on scalability, simplicity, and accessibility.
đ Quick Start
Data
We've released all our curated high - quality training data in the data
folder:
- Curated 129k data:
- Original 57k, collected from various sources such as AIME (up to 2023), MATH, Numina - Math collection, and Tulu3 MATH.
- Extended 72k, mainly cleaned from OpenR1 - Math - 220k.
- Hard 13k, mined from the first stage of ORZ - 32B training.
For details on data collection, refer to our paper.
Installation & Training Scripts
We've provided a Dockerfile in the docker folder to ensure reproducibility of our training.
To install the package, run:
pip install -e .
Start ORZ - 32B PPO Training
Here are the commands to start training on 16 nodes:
- On the master node, run:
ray start --head
# You'll see logging like:
# Next steps
# To add another node to this Ray cluster, run
# ray start --address='<master-node-ip>:<master-node-port>'
- On all other nodes, run:
ray start --address='<master-node-ip>:<master-node-port>' # <master-node-ip> and <master-node-port> are from above loggings!
- Finally, on the master node, run:
python -m playground.orz_32b_ppo
The training log will be displayed in the master node terminal.
Start ORZ - 0.5B PPO Training
You can start ORZ - 0.5B PPO training on a single A800/H800 node:
python -m playground.orz_0p5b_ppo
You can even run it on a single A800/H800 GPU:
python -m playground.orz_0p5b_ppo_1gpu
Note: Since it's not a multi - node setting, there's no need for ray start
commands.
Start ORZ - 7B PPO Training
For multi - node training on 4 nodes:
# Set up for multi - node training
ray start --head # on master node
ray start --address='<master-node-ip>:<master-node-port>' # then on other nodes
# Then on the master node, run:
python -m playground.orz_7b_ppo
The training log will be shown in the master node terminal.
Start ORZ - 1.5B PPO Training
For multi - node training on 2 nodes:
# Set up for multi - node training
ray start --head # on master node
ray start --address='<master-node-ip>:<master-node-port>' # then on other nodes
# Then on the master node, run:
python -m playground.orz_1p5b_ppo
Debug Settings
In the code, we've left an environment variable DEBUG_MODE
for researchers to run in debug mode. (Currently, we recommend using python -m playground.orz_0p5b_ppo_1gpu
for debugging.)
Here are some debug running command examples:
# NOTE: just for debug, not final setting!
## Debug command in a single GPU with `EleutherAI/pythia-14m`
DEBUG_MODE=True python -m playground.orz_14m_ppo_mini
## Debug command in a single node (8 GPUs) with `Qwen/Qwen2.5-7B`
DEBUG_MODE=True python -m playground.orz_7b_ppo
How to Use the Model
Policy Model
Policy models can be used in the same way as any chat model in transformers
and vllm
, as we've included the chat template jinja in the tokenizer.
Critic Model
Critic models can be loaded in the same way as in the training code.
⨠Features
- Single Controller Trainer Design: Adopts a single controller trainer design, which is flexible and researcher - friendly.
- Maximized GPU Utilization: Colocates training and generation on the same GPUs to maximize GPU utilization.
đĻ Installation
We've provided a Dockerfile in the docker folder. To install the package, run:
pip install -e .
đ Documentation
Overview đ
We introduce Open - Reasoner - Zero, the first open - source implementation of large - scale reasoning - oriented RL training, emphasizing scalability, simplicity, and accessibility.
Using the same base model as DeepSeek - R1 - Zero - Qwen - 32B, our implementation achieves superior performance on AIME2024, MATH500, and the GPQA Diamond benchmark. It also demonstrates remarkable efficiency, requiring only a tenth of the training steps compared to the DeepSeek - R1 - Zero pipeline.
To encourage broader participation and accelerate research towards artificial general intelligence (AGI), we've released our source code, parameter settings, training data, and model weights. For more insights across various model sizes, please refer to our paper.
Let the Reasoner - Zero tide rise!
Main Results đ
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.
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.
Releases đĻ
[2025/03/31]
We've reached a major milestone for Open - Reasoner - Zero
:
- đ Updated Paper with new results.
- đ Easy - to - use Training Scripts:
- ORZ - 1.5B training scripts and ORZ - 0.5B training scripts (main results in Figure 2).
- Minimal resource training scripts: ORZ - 0.5B can be run on a single A800/H800 GPU!
- 𤊠Updated Curated Datasets:
- 129k data in total:
- 13k hard data mined from the above 129k data, used in the "annealing" stage of ORZ - 32B training, boosting AIME2024 from ~41% to ~48%!
- đ¤ More HF Models:
- Updated HF Models:
Open - Reasoner - Zero - 7B
andOpen - Reasoner - Zero - 32B
. - Released HF Models:
Open - Reasoner - Zero - 1.5B
andOpen - Reasoner - Zero - 0.5B
.
- Updated HF Models:
- đ Full Suite of Critic Models for in - depth research:
Open - Reasoner - Zero - Critic -
{0.5B, 1.5B, 7B, 32B}.
[2025/02/18]
We released Open - Reasoner - Zero
. As part of this release, we open - sourced:
- đ Paper on our comprehensive analysis and insights in Reasoner - Zero training.
- đ¤ HF Model
Open - Reasoner - Zero - 7B
andOpen - Reasoner - Zero - 32B
. - đ
Our curated 57k training data
. - đ Training Scripts for you to start your own Reasoner - Zero journey!
Acknowledgements đ
- This work was supported by the computing resources and valuable feedback from StepFun and Tsinghua University.
- Our training framework is built on OpenRLHF, vllm, DeepSpeed, and ray.
- Our model is based on the Qwen2.5 Series of base models, including Qwen2.5 - 0.5B, Qwen2.5 - 1.5B, Qwen2.5 - 7B, and Qwen2.5 - 32B.
- We thank Project Numina, Tulu3, and OpenR1 - Math - 220k for their collected open - sourced data.
Advertisement Time đŖ
We're hiring talented researchers and engineers to join our team. If you're interested in our project and want to contribute to the reasoner scale - up towards AGI, please contact us at hanqer@stepfun.com
[](https://star - history.com/#Open-Reasoner-Zero/Open-Reasoner-Zero&Timeline)
Community Discussions đē
We have several WeChat groups for discussions and sharing. You can scan the QR code below to join the latest group.
đ License
The project is licensed under the MIT license.
Citation
@misc{hu2025openreasonerzeroopensourceapproach,
title={Open-Reasoner-Zero: An Open Source Approach to Scaling Up Reinforcement Learning on the Base Model},
author={Jingcheng Hu and Yinmin Zhang and Qi Han and Daxin Jiang and Xiangyu Zhang and Heung-Yeung Shum},
year={2025},
eprint={2503.24290},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2503.24290},
}

