Open-SWE-Traces: Advancing Dual-Mode Multilingual Distillation for Software Engineering Agents

A research team has released Open-SWE-Traces, a dataset of 207,489 agentic trajectories designed to train multilingual software engineering models, according to a paper published this week [1]. The dataset spans nine programming languages — Python, Go, TypeScript, JavaScript, Rust, Java, PHP, C, and C++ — and was sourced from 20,000 real-world pull requests using the OpenHands and SWE-agent harnesses [1]. The trajectories were generated through a hybrid-reasoning synthesis: Minimax-M2.5 produced traces with explicit reasoning steps, while Qwen3.5-122B contributed traces without such processes [1]. All data was filtered for permissive licenses, including MIT, Apache, and BSD, drawn from the SWE-rebench-V2 benchmark [1]. To validate the resource, the authors fine-tuned models from the Qwen3-30B-A3B series, covering Thinking, Instruct, and Coder variants [1]. The top-performing model achieved a resolve rate of 61.7% on SWE-bench Verified, 57.1% on SWE-bench Multilingual, and 36.8% on SWE-bench Pro [1]. These benchmarks test a system’s ability to autonomously resolve real-world GitHub issues. The paper, authored by Wasi Uddin Ahmad and colleagues, argues that a deficit of diverse, large-scale trajectory data has constrained progress toward autonomous software engineering [1]. The release adds to a growing body of work examining how datasets can complement one another for training machine learning models. Prior studies in other scientific domains, such as computational catalysis, have explored whether data from different sources is complementary and how transfer learning can improve performance when consolidating varied computational methods [4]. Those efforts have shown that models trained jointly on multiple datasets or fine-tuned via transfer learning can outperform those trained on a single source [4]. The Open-SWE-Traces team positions its dataset as a resource for distilling human-level software engineering capabilities into efficient, open-source agentic large language models [1].