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

2026-06-14

AI Summary

Current autonomous software engineering is hindered by the scarcity of high-quality, multilingual agent trajectory data. This work addresses this limitation by constructing a large-scale multilingual dataset comprising 207,489 trajectories across nine programming languages, derived from 20,000 real-world pull requests. The study further introduces a novel dual-modality trajectory distillation framework—supporting both explicit and implicit reasoning—and leverages trajectories generated by Minimax-M2.5 and Qwen3.5-122B, collected via OpenHands and SWE-agent. Fine-tuning on the Qwen3-30B-A3B model series yields significant performance gains, with the best-performing model achieving problem-resolution rates of 61.7%, 57.1%, and 36.8% on SWE-bench Verified, Multilingual, and Pro benchmarks, respectively, substantially advancing the capabilities of open-source agents in multilingual software engineering tasks.

Abstract

The path toward autonomous software engineering is currently bottlenecked by a severe deficit of diverse, large-scale trajectory data. We address this by introducing \ourdataset, an expansive dataset of 207,489 agentic trajectories spanning nine programming languages (Python, Go, TS, JS, Rust, Java, PHP, C, C++). Sourced from 20,000 real-world PRs via OpenHands and SWE-agent harnesses, the dataset utilizes a hybrid-reasoning synthesis: Minimax-M2.5 generates trajectories with explicit "thinking" processes, while Qwen3.5-122B provides high-quality "non-thinking" traces. Filtered for permissive licenses (MIT, Apache, BSD) from SWE-rebench-V2, this data facilitates the training of models capable of long-horizon reasoning. We validate the dataset by fine-tuning the Qwen3-30B-A3B series (Thinking, Instruct, and Coder). The best performing model achieves resolve rates of 61.7% on SWE-bench Verified, 57.1% on SWE-bench Multilingual, and 36.8% on SWE-bench Pro. These results establish Open-SWE-Traces as a premier resource for distilling human-level software engineering capabilities into efficient, open-source agentic LLMs.

Problem

• autonomous software engineering
• trajectory data
• multilingual
• agentic LLMs
• data deficit

Innovation

• dual-mode distillation
• agentic trajectories
• multilingual software engineering
• long-horizon reasoning
• open-source LLMs