SWE-Dev-7B: An Open-Source Agent for Software Engineering

SWE-Dev-7B, developed by THUDM, is a 7.6 billion parameter model built upon the Qwen-2.5-Coder-7B-Instruct architecture. It is part of the SWE-Dev family of models, which are open-source agents specifically engineered for software development tasks. The project introduces a novel pipeline for generating developer-centric datasets from GitHub, encompassing issue tracking, code localization, and test case generation.

Key Capabilities

• Software Engineering Tasks: Optimized for a range of software development activities, including understanding and resolving issues, locating relevant code sections, and generating test cases.
• Performance on SWE-bench: SWE-Dev-7B achieved a 23.4% solve rate on SWE-bench-Verified, demonstrating strong performance in automated software engineering.
• Training Data Scaling: Research indicates that scaling training data and improving data quality, especially when combined with reinforcement fine-tuning (RFT), significantly boosts performance on SWE-bench.
• Inference Scaling: The model's performance also benefits from inference scaling, with solve rates increasing with more inference rounds.

Good For

• Automated software development and bug fixing.
• Code generation and understanding in a software engineering context.
• Researchers and developers looking for models specialized in handling GitHub repository data and software issues.