Beijing, China – April 13, 2024 – Kunlun Wanwei, a leading Chinese technology company, today announced the release of the Skywork-OR1 (Open Reasoner 1) series of large language models (LLMs). This new series, an evolution of their initial Chinese logical reasoning model, Skywork-o1, boasts industry-leading reasoning performance within comparable parameter scales and is poised to significantly advance the capabilities of LLMs in logical understanding and complex task-solving.
In a move that underscores their commitment to open innovation, Kunlun Wanwei is making the Skywork-OR1 series fully accessible, free to use, and completely open-source, aiming to empower the developer community. The release encompasses three high-performance models:
- Skywork-OR1-Math-7B: A specialized model focusing on mathematical domains, while also exhibiting strong coding abilities.
- Skywork-OR1-7B-Preview: A general-purpose model blending mathematical and coding capabilities, offering versatility and specialization.
- Skywork-OR1-32B-Preview: A flagship version designed for tackling more complex tasks, equipped with enhanced reasoning capabilities.
We believe in the power of open collaboration and its potential to accelerate innovation in the field of artificial intelligence, stated a Kunlun Wanwei spokesperson. By releasing the Skywork-OR1 series under a fully open-source license, we hope to foster a vibrant community of developers and researchers who can build upon our work and push the boundaries of what’s possible.
Unprecedented Transparency and Reproducibility
Kunlun Wanwei is adopting the highest level of transparency with the Skywork-OR1 series, open-sourcing not only the model weights but also the training datasets and complete training code. All resources have been uploaded to GitHub and Hugging Face platforms. A detailed technical blog, available on Notion, elaborates on the data processing pipeline, training methodologies, and key technical insights, providing the community with a fully reproducible practical reference.
The open-source nature of the Skywork-OR1 series allows researchers and developers to:
- Understand the inner workings of the models: Gain insights into the model architecture, training process, and data used.
- Reproduce the results: Verify the performance claims and conduct independent evaluations.
- Customize and fine-tune the models: Adapt the models to specific applications and datasets.
- Contribute to the development of the models: Collaborate with the community to improve the models and address potential issues.
Redefining Evaluation Metrics: Introducing avg@k
The Skywork-OR1 series introduces avg@k as a core evaluation metric, designed to measure the average performance of the model in successfully solving a problem across k attempts. This metric offers a more nuanced understanding of the model’s stability and overall reasoning ability compared to the traditional pass@k metric, which only focuses on whether the model succeeds at least once. avg@k provides a more comprehensive reflection of the model’s true performance level and practical value.
Exceptional Performance in Mathematical Reasoning
The Skywork-OR1 series demonstrates exceptional performance in mathematical reasoning tasks. The general-purpose models, Skywork-OR1-7B-Preview and Skywork-OR1-32B-Preview, achieved optimal performance within their parameter scale on the AIME24 and AIME25 datasets, showcasing their robust mathematical reasoning capabilities. The specialized model, Skywork-OR1-Math-7B, further excels on the AIME dataset, highlighting the benefits of targeted optimization for specific domains.
Looking Ahead
Kunlun Wanwei plans to release official versions of Skywork-OR1-7B and Skywork-OR1-32B within the next two weeks, along with a more comprehensive technical report detailing their experiences and insights in training reasoning models. This ongoing development and commitment to sharing knowledge solidify Kunlun Wanwei’s position as a leader in open-source AI.
Open-Source Resources:
- Skywork-OR1 Series: https://github.com/SkyworkAI/Skywork-OR1
- Kunlun Wanwei TianGong Team Open Source Projects: https://huggingface.co/Skywork
Conclusion
Kunlun Wanwei’s release of the Skywork-OR1 series marks a significant step forward in the development of open-source large language models. By providing fully transparent and accessible models, Kunlun Wanwei is empowering the AI community to innovate and build upon their work. The Skywork-OR1 series, with its exceptional reasoning performance and commitment to open-source principles, is poised to drive advancements in various fields, from mathematics and coding to complex problem-solving. Future research could focus on further refining the models, exploring new applications, and expanding the open-source ecosystem around the Skywork-OR1 series. This release underscores the growing importance of open collaboration in shaping the future of artificial intelligence.
Views: 0