SWE-bench/SWE-smith
A toolkit for training and scaling data for SWE-agents that can turn GitHub repos into executable training environments and train LMs to become better software engineers.
Velocity · 7d
+1.7
★ / day
Trend
→steady
star history
SWE-smith provides tooling to build unlimited training tasks for software engineering agents by creating Docker-based execution environments from GitHub repositories. It synthesizes task instances, generates harnesses that break unit tests, and produces issue text for training data. The toolkit also trains language models like SWE-agent-LM-32B to improve agent performance on software engineering tasks, appearing at NeurIPS 2025 Datasets & Benchmarks Track.