Skywork-OR1-7B-Preview: An Open Reasoner for Math and Code
Skywork-OR1-7B-Preview is a 7.6 billion parameter model from the Skywork-OR1 (Open Reasoner 1) series, developed by Skywork. This model is designed as a general-purpose reasoner with a strong focus on mathematical and coding tasks. It is trained using a sophisticated pipeline involving large-scale rule-based reinforcement learning (RL) and a customized version of GRPO.
Key Capabilities & Features
- Enhanced Reasoning: Optimized for complex mathematical problems and coding challenges.
- Performance: Outperforms other models of similar size in both math and coding benchmarks, as measured by Avg@K metrics on AIME24, AIME25, and LiveCodeBench.
- Advanced Training: Utilizes a multi-stage training pipeline with adaptive entropy control, difficulty-based filtering, and rejection sampling for improved efficiency and stability.
- Curated Data: Trained on a meticulously selected and cleaned dataset of 110K math problems and 14K coding questions, with model-aware difficulty estimation.
When to Use This Model
- Mathematical Problem Solving: Ideal for applications requiring strong mathematical reasoning, including competitive programming or scientific calculations.
- Code Generation & Analysis: Suitable for tasks involving code understanding, generation, and debugging, demonstrated by its performance on LiveCodeBench.
- Research in Reasoning: Provides an open-source foundation for further research into reasoning models, with detailed training recipes and data available.