Skywork-SWE-32B: A Code Agent for Software Engineering
Skywork-SWE-32B, developed by Skywork AI, is a 32.8 billion parameter code agent model built upon the Qwen2.5-Coder-32B-Instruct backbone. It is specifically engineered for software engineering (SWE) tasks, demonstrating strong performance in automated code generation and bug fixing.
Key Capabilities & Performance
- SWE-bench Verified Accuracy: Achieves 38.0% pass@1 accuracy on the SWE-bench Verified benchmark, surpassing previous open-source Qwen2.5-Coder-32B-based LLMs when integrated with the OpenHands agent framework.
- Enhanced Performance with Test-Time Scaling: Performance further improves to 47.0% accuracy when incorporating test-time scaling techniques, setting a new state-of-the-art for sub-32B parameter models.
- Data Scaling Law Demonstration: The model's development clearly illustrates the data scaling law phenomenon for software engineering capabilities in LLMs, indicating continued performance gains with more training data.
- High-Quality Training Data: Trained on the Skywork-SWE dataset, a large-scale, high-quality collection featuring comprehensive executable runtime environments, gathered via an efficient and automated pipeline.
Recommended Use Cases
- Automated Software Development: Ideal for tasks requiring automated code generation, bug fixing, and patch creation within software projects.
- Integration with Agent Frameworks: Designed to work effectively with agent frameworks like OpenHands for complex software engineering workflows.
- Research on LLM Scaling Laws: Useful for researchers studying the impact of data scaling on LLM performance in specialized domains like software engineering.