Open-SWE-Traces is a dataset released by NVIDIA in 2026 for fine-tuning instructions for large language model agents. Related papers include... OPEN-SWE-TRACES: Advancing Dual-Mode Multilingual Distillation for Software Engineering AgentsIt aims to improve the model's ability to repair code and call multi-step tools in the field of software engineering. It is widely used in code agent training, automated software repair system development, SWE-Bench benchmark optimization, and supervised fine-tuning and distillation research of large models in software engineering tasks. This dataset contains 207,489 agent interaction trajectories, synthesized using the Minimax-M2.5 and Qwen3.5-122B-A10B algorithms, collected using the SWE-agent and OpenHands framework, covering multiple programming languages including Python, Go, TypeScript, JavaScript, Rust, Java, PHP, C, and C++, as well as the performance distribution of these Agent-Models across different programming languages.

Dataset composition:

• OpenHands + Minimax-M2.5: 49,948 tracks; 20,362 PRs (Pull Requests/Number of Tasks); PR Status (Resolved / Unresolved / Unknown): 15,941 / 22,911 / 11,096
• OpenHands + Qwen3.5-122B: 55,488 tracks; 20,098 PRs (Pull Requests/Number of Tasks); PR Status (Resolved / Unresolved / Unknown): 13,657 / 29,147 / 12,684
• SWE-agent + Minimax-M2.5: 57,268 tracks; 20,791 PRs (Pull Requests/Number of Tasks); PR Status (Resolved / Unresolved / Unknown): 19,551 / 24,554 / 13,163
• SWE-agent + Qwen3.5-122B: 44,785 tracks; 18,211 PRs (Pull Requests/Number of Tasks); PR Status (Resolved / Unresolved / Unknown): 16,095 / 18,875 / 9,815

Citation

@article{ahmad2026openswetraces,
title={OPEN-SWE-TRACES: Advancing Dual-Mode Multilingual Distillation for Software Engineering Agents},
author={Wasi Uddin Ahmad, Nikolai Ludwig, Somshubra Majumdar, Boris Ginsburg},
year={2026},
eprint={2606.16038},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2606.16038},
}