Autonomous software engineering agents have long been limited by a lack of diverse, large-scale trajectory data. A new dataset, Open-SWE-Traces, aims to close that gap by providing 207,489 agentic trajectories across nine programming languages, according to a paper by Ahmad, Ludwig, Majumdar, and Ginsburg posted on arXiv.
Dataset Composition and Synthesis
The trajectories were sourced from 20,000 real-world pull requests (PRs) using the OpenHands and SWE-agent harnesses. They cover Python, Go, TypeScript, JavaScript, Rust, Java, PHP, C, and C++. To generate the trajectories, the authors employed a hybrid-reasoning synthesis: Minimax-M2.5 produced trajectories with explicit "thinking" processes, while Qwen3.5-122B generated high-quality "non-thinking" traces. All data was filtered to include only permissive licenses (MIT, Apache, BSD) from SWE-rebench-V2.
Validation and Performance
The dataset's effectiveness was validated by fine-tuning the Qwen3-30B-A3B series (Thinking, Instruct, and Coder). The best performing model achieved the following resolve rates on standard benchmarks:
| Benchmark | Resolve Rate |
|---|---|
| SWE-bench Verified | 61.7% |
| SWE-bench Multilingual | 57.1% |
| SWE-bench Pro | 36.8% |
These results establish Open-SWE-Traces as a premier resource for distilling human-level software engineering capabilities into efficient, open-source agentic LLMs, according to the paper.
Implications for Enterprise Software Engineering
For CTOs and technology procurement leaders, the availability of open-source, multilingual agentic trajectory data could reduce the cost and complexity of building custom software engineering agents. Fine-tuning on Open-SWE-Traces allows models to perform long-horizon reasoning across codebases in languages from Python to C++, potentially accelerating development pipelines. The dataset, derived from real PRs, reflects genuine human coding practices rather than synthetic scenarios.
The use of both thinking and non-thinking trajectories gives developers flexibility: thinking traces can be used for interpretability, while non-thinking traces optimize inference speed. As autonomous agents move toward production use, such resources lower the barrier to entry for enterprises seeking to automate parts of their software lifecycle.
Availability and Licensing
Open-SWE-Traces is released under permissive licenses (MIT, Apache, BSD), making it suitable for commercial use. The full dataset and fine-tuning recipes are available via the paper's accompanying code repository, enabling organizations to reproduce and extend the results.