Skywork-OR1

🔥 News

• May 29, 2025: Our Skywork Open Reasoner 1 Technical Report has been released on arXiv. We provide further details on the training pipeline, investigation and mitigation to the entropy collapse phenomenon, and extensive analysis and ablation studies.
• May 13, 2025: We release our final version of Skywork-OR1 series of models：Skywork-OR1-32B and Skywork-OR1-7B.
  • Skywork-OR1-32B outperforms Deepseek-R1 and Qwen3-32B on math tasks (AIME24 and AIME25) and delivers comparable performance on coding tasks (LiveCodeBench).
  • Skywork-OR1-7B exhibits competitive performance compared to similarly sized models in both math and coding scenarios.
• April 15, 2025: We release our rl training dataset Skywork-OR1-RL-Data.
• April 13, 2025: We release the Skywork-OR1 (Open Reasoner 1) series of models, including Skywork-OR1-Math-7B, Skywork-OR1-32B-Preview, and Skywork-OR1-7B-Preview. We open-source
  • 🤗 Model weights: Skywork-OR1-Math-7B, Skywork-OR1-32B-Preview, Skywork-OR1-7B-Preview
  • 🤗 Training data: Skywork-OR1-RL-Data
  • 🧑‍💻 Code: Skywork-OR1
  • We also release a Notion Blog to share detailed training recipes and extensive experimental results, analysis, and insights, dedicated to helping the community to better research, understand, and push the frontier of open reasoning models.

📖 Overview

_{The AIME24 and AIME225 scores versus training steps of Skywork-OR1-32B in our multi-stage training pipeline.}

The Skywork-OR1 (Open Reasoner 1) model series consists of powerful math and code reasoning models trained using large-scale rule-based reinforcement learning with carefully designed datasets and training recipes. This series includes two general-purpose reasoning modelsl, Skywork-OR1-7B and Skywork-OR1-32B.

• Skywork-OR1-32B outperforms Deepseek-R1 and Qwen3-32B on math tasks (AIME24 and AIME25) and delivers comparable performance on coding tasks (LiveCodeBench).
• Skywork-OR1-7B exhibits competitive performance compared to similarly sized models in both math and coding scenarios.

📊 Evaluation

We evaluate our models on AIME24, AIME25, and LiveCodeBench. Instead of using Pass@1, which is common in prior work, we introduce Avg@K as the primary metric. This metric robustly measures a model’s average performance across K independent attempts, reducing the impact of randomness and enhancing the reliability of the results. We believe that Avg@K provides a better reflection of a model’s stability and reasoning consistency.

We include the detailed results in the following table.

🎯 Getting Started

Installation

Docker environment:

docker pull whatcanyousee/verl:vemlp-th2.4.0-cu124-vllm0.6.3-ray2.10-te2.0-megatron0.11.0-v0.0.6

# Launch the desired Docker image:
docker run --runtime=nvidia -it --rm --shm-size="10g" --cap-add=SYS_ADMIN -v <image:tag>

# Inside the container, install Skywork-OR1
git clone https://github.com/SkyworkAI/Skywork-OR1.git && cd Skywork-OR1 && pip3 install -e .

Conda environment:

# Installing Python 3.10 Environment.
conda create -n verl python==3.10
conda activate verl

# Installing RLLM dependencies.
pip3 install torch==2.4.0 --index-url https://download.pytorch.org/whl/cu124
pip3 install flash-attn --no-build-isolation
git clone https://github.com/SkyworkAI/Skywork-OR1.git
cd Skywork-OR1
pip3 install -e .

Training ⚙️

We provide training scripts and data to reproduce the results of the “Skywork-OR1-Series”.

Training Data Preparation

To prepare the training data, we provide a script to download the data from Hugging Face and filter the problems based on the difficulty level with respect to a particular model (i.e., DeepSeek-R1-Distill-Qwen-{1.5,7,32}B).

model_size=32b  # 1p5b, 7b
python ./or1_scripts/data_preprocess/download_and_filter_data_${model_size}.py --local_dir ./or1_data/train

This will generate the training data in the following format:

./or1_data/train/train_${model_size}_math.pkl
./or1_data/train/train_${model_size}_code.pkl

Train Script

By default, we only provide evaluation on AIME datasets. If you would like to evaluate on LiveCodeBench, please refer to the section Evaluation Data Preparation and set LIVECODEBENCH_DATA_PATH to ./or1_data/eval/livecodebench/livecodebench_2408_2502.

# Note: You must provide CODE_PATH and MODEL_PATH
model_size=7b # or 32b
train_seq_len=8 # or 16, 32
export CODE_PATH=./
export MODEL_PATH=
bash ./or1_scripts/train/${model_size}_${train_seq_len}k.sh

Using Ray for Multi-Node Training

If you plan to perform multi-node training, you need to start and connect all nodes using Ray before launching the training script. Here’s a quick guide to set up Ray across machines:

Step 1: Start Ray on the Head Node (node0)

On the first node (typically called node0), run:

ray start --head --dashboard-host=0.0.0.0

After running the command, you will see a message like:

Ray runtime started.
Next steps
To add another node to this Ray cluster, run
    ray start --address='10.94.16.4:6379'

Note down the IP address (in this example, 10.94.16.4).

Step 2: Connect Other Nodes (e.g., node1)

On each additional worker node (e.g., node1), run the following, replacing the IP with that of your head node:

ray start --address='10.94.16.4:6379'

Step 3: Check Cluster Status

On node0, run:

ray status

You should see output showing all connected nodes and available resources (e.g., CPUs, GPUs, memory). For example:

Resources
---------------------------------------------------------------
Usage:
 0.0/360.0 CPU
 0.0/16.0 GPU
...

Once the Ray cluster is up and running, you can launch the training script as usual. The script will automatically utilize the connected nodes.

Evaluation ⚖️

We provide evaluation scripts to reproduce the results of the Skywork-OR1-Series.

Evaluation Data Preparation

Evaluation data for AIME24 and AIME25 is already available in our GitHub repository.

For LiveCodeBench, please download the data from Hugging Face.

# Download LiveCodeBench
huggingface-cli download Skywork/LiveCodeBench --repo-type=dataset --local-dir ./or1_data/eval/livecodebench
unzip ./or1_data/eval/livecodebench/livecodebench.zip -d ./or1_data/eval/livecodebench/
mv ./or1_data/eval/livecodebench/livecodebench/* ./or1_data/eval/livecodebench/

Evaluation Start

bash ./or1_scripts/eval/eval_7b.sh

bash ./or1_scripts/eval/eval_32b.sh

The evaluation results will be automatically saved to outputs/evalation/pass.csv

📄 Technical Report

Our technical report will be released soon. Stay tuned!

🙏 Acknowledgements

• Both of our models are trained on top of DeepSeek-R1-Distill-Qwen-7B and DeepSeek-R1-Distill-Qwen-32B.
• Both models are trained using a custom fork of the wonderful verl project.

📚 Citation

Please cite the following:

@article{he2025skywork,
  title={Skywork Open Reasoner 1 Technical Report},
  author={He, Jujie and Liu, Jiacai and Liu, Chris Yuhao and Yan, Rui and Wang, Chaojie and Cheng, Peng and Zhang, Xiaoyu and Zhang, Fuxiang and Xu, Jiacheng and Shen, Wei and Li, Siyuan and Zeng, Liang and Wei, Tianwen and Cheng, Cheng and An, Bo and Liu, Yang and Zhou, Yahui},
  journal={arXiv preprint arXiv:2505.22312},
  year={2025}
}

@misc{skywork-or1-2025,
  title={Skywork Open Reasoner Series},
  author = {He, Jujie and Liu, Jiacai and Liu, Chris Yuhao and Yan, Rui and Wang, Chaojie and Cheng, Peng and Zhang, Xiaoyu and Zhang, Fuxiang and Xu, Jiacheng and Shen, Wei and Li, Siyuan and Zeng, Liang and Wei, Tianwen and Cheng, Cheng and An, Bo and Liu, Yang and Zhou, Yahui},
  howpublished={\url{https://capricious-hydrogen-41c.notion.site/Skywork-Open-Reaonser-Series-1d0bc9ae823a80459b46c149e4f51680}},
  note={Notion Blog},
  year={2025}
}