princeton-nlp/SWE-Llama-7b
SWE-Llama-7b is a 7 billion parameter CodeLlama-based transformer model developed by Princeton NLP, specifically fine-tuned for software engineering tasks. It excels at generating patches to resolve real-world GitHub issues, conditioned on issue descriptions and code context. This model is optimized for automated code repair and issue resolution, distinguishing it from general-purpose code generation models.
Loading preview...
SWE-Llama-7b: CodeLlama for Software Engineering
SWE-Llama-7b is a 7 billion parameter model from Princeton NLP, built upon the CodeLlama architecture. It is uniquely fine-tuned for software engineering tasks, focusing on the automated resolution of real-world GitHub issues and pull requests.
Key Capabilities
- Automated Issue Resolution: Designed to generate code patches that address GitHub issues, leveraging issue descriptions and surrounding code context.
- Specialized Training: Fine-tuned on a dataset of 19,000 issues and pull requests from 37 popular Python repositories, distinct from the SWE-bench evaluation set.
- Efficient Fine-tuning: Utilizes the LoRA method to fine-tune only the attention matrices, trained for 4 epochs.
Performance
On the SWE-bench benchmark, SWE-Llama-7b achieved a 3.0% issue resolution rate using oracle context retrieval, demonstrating its capability in practical software repair scenarios.
Good For
- Automated Code Repair: Ideal for systems requiring automated generation of fixes for identified software bugs or issues.
- Developer Tooling: Can be integrated into developer workflows to assist with patch generation and issue resolution.
- Research in Software Engineering: A valuable base model for further research into AI-driven software development and maintenance.