SWE-smith

A synthetic SWE-task generation pipeline (50K+ instances) that trained SWE-agent-LM-32B, the best open-weights SWE agent at release, on SWE-bench tasks. NeurIPS 2025 Datasets & Benchmarks Spotlight; later expanded beyond Python to Go, Java, JavaScript, Rust, C++, and TypeScript. Produced by Princeton (with Stanford collaborators) as part of the SWE-agent ecosystem. The trained model is fine-tuned from an open base, not pretrained from scratch.