OpenSWE
- Type
- dataset
- Venue
- GAIR / SII-GAIR-NLP (HuggingFace)
- Year
- 2026
- Source
- huggingface
- Access
- restricted
- Language
- English, Python
- Added
- 2026-07-17T20:18:03.753806+00:00
- Verified
- 2026-07-17T20:18:03.753806+00:00
Summary
OpenSWE is the largest fully transparent framework for training software engineering (SWE) agents, comprising 45,320 executable Docker environments across 12.8k repositories with all Dockerfiles, evaluation scripts, and infrastructure open-sourced for reproducibility. Built via a multi-agent synthesis pipeline on a 64-node distributed cluster, it yields ~13,000 curated trajectories from ~9,000 quality-guaranteed environments and produces models (OpenSWE-32B/72B) achieving SOTA among SFT-based methods on SWE-bench Verified (62.4%/66.0%).
Keywords
software-engineering swe-agent code pull-request docker synthetic agent swe-bench
Topics
Code / Software Engineering
Research notes
- Requires accepting conditions and sharing contact info to access. Construction cost ~$891K (environments) + $576K (trajectory sampling/curation). Associated arXiv papers: 2603.13023, 2505.20411, 2506.10954.