Open-SWE-Traces: Advancing Distillation for Software Engineering Agents

🚨 What's New

• [09/26] v1.2 Release: Added new agent trajectories generated by Qwen3.8-27B for mini-SWE-agent and OpenCode (coming soon) harnesses.
• [08/26] v1.1 Release: Added new agent trajectories generated by DeepSeek-V4-Flash and Qwen3.6-27B across OpenHands, SWE-agent, and mini-SWE-agent harnesses.
• [06/21] v1.0 Release: Released 207k agent trajectories across OpenHands and SWE-agent harnesses. On 08/26, we removed trajectories that include git hacking behaviors (detected using an LLM-as-a-judge). Now, v1.0 includes 151k agent trajectories.

Data Overview

Open-SWE-Traces is an agentic instruction tuning dataset designed to advance the capabilities of LLMs in software engineering. This dataset comprises 200k+ agent trajectories collected using the SWE-agent and OpenHands framework. The trajectories were synthesized using Minimax-M2.5 (with thinking) and Qwen3.5-122B-A10B (without thinking) and specifically curated for supervised fine-tuning (SFT), aiming to improve model performance on SWE-Bench style tasks. The issue statements are sourced from SWE-rebench-V2 with permissive licenses (MIT, Apache-2.0, BSD-2-Clause, BSD-3-Clause).

This dataset is ready for commercial/non-commercial use.

Data Fields

Data distribution (before git hacking behavior filtering)

The Open-SWE-Traces dataset consists of 207,489 total trajectories across two LLMs and agent scaffolds.

Table 1: Agent, Model, and PR Category-wise Distribution

Table 2: Language-wise Distribution

Note:

• In Table 1, Resolved, Unresolved, and Unknown indicate issue resolution status.
• In Table 2, The counts are formatted as Trajectories / PRs.
  • PR counts are non-unique across categories and cannot be summed horizontally.

Dataset Characterization

** Data Collection Method

• [Hybrid: Automated, Synthetic]
  

** Labeling Method

• [Hybrid: Automated, Synthetic]
  

License/Terms of Use

This dataset is governed by the Creative Commons Attribution 4.0 International License (CC BY 4.0). Additional Information: MIT License, Apache 2.0 License, BSD 2 Clause License and BSD 3 Clause License.

Data Developer: NVIDIA

Data Version

1.0

Intended use

This dataset is intended for LLM engineers and research teams building autonomous software engineering agents and code-focused assistants. It is suitable for supervised fine-tuning and distillation of models that must interpret real-world issue statements, plan multi-step tool use, navigate codebases, and implement fixes in a SWE-Bench–style setting. The trajectories can also be used to benchmark and debug agent policies, improve repository-aware reasoning, and study robust, regression-free code editing behaviors in both academic and production environments.

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report security vulnerabilities or NVIDIA AI Concerns here.

Citation

If you find the data useful, please cite:

@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}, 
}