alucent/mirror-SWE-Next-SFT-Trajectories

SWE-Next: Scalable Real-World Software Engineering Tasks for Agents

SWE-Next SFT Trajectories

SWE-Next SFT Trajectories is the supervised fine-tuning dataset released with SWE-Next: Scalable Real-World Software Engineering Tasks for Agents. It contains 3,693 ShareGPT-style multi-turn training examples collected from expert agent rollouts on 2,308 execution-grounded SWE tasks synthesized from real merged pull requests.

The dataset is designed for training repository-level SWE agents rather than isolated code generators. Each example is a cleaned interaction trace with roles such as system, user, assistant, and tool, suitable for direct use in LLaMA-Factory and similar chat-style SFT pipelines.

Dataset Overview

SWE-Next first mines real merged PRs, executes candidate base/merged commit pairs, and retains only commit pairs that yield strict test improvements without regressions. Expert models then interact with these self-verifying tasks under a fixed repository-level interface. From the resulting rollouts, we keep two buckets for SFT:

• Clean successes: final verification passes, a real code edit is made, and earlier test evidence remains consistently parseable.
• Recovery successes: final verification passes after at least one earlier failing test, capturing repair trajectories with meaningful debugging evidence.

This filtering yields a compact but high-signal SFT corpus for downstream SWE-agent training.

Format

The dataset contains one training split with 3,693 rows.

Each row is a JSON object of the form:

{"messages": [{"role": "system", "content": "..."}, {"role": "user", "content": "..."}, ...]}

The messages field follows a ShareGPT-style schema:

• system: agent setup and high-level instructions
• user: task statement and repository-level problem context
• assistant: the model's intermediate reasoning/actions
• tool: tool outputs such as shell execution, file inspection, and test feedback

Files

• SWE_Next_SFT_Trajectories.jsonl: the released ShareGPT-style SFT corpus

Usage

Load the dataset with Hugging Face Datasets:

from datasets import load_dataset

ds = load_dataset("TIGER-Lab/SWE-Next-SFT-Trajectories")
print(ds["train"][0].keys())

It can also be used directly with LLaMA-Factory-style training configs by pointing the training pipeline to SWE_Next_SFT_Trajectories.jsonl.

Relationship to the SWE-Next Release

This repo contains only the released SFT trajectories. Related artifacts are available separately:

• Base task dataset: TIGER-Lab/SWE-Next
• Released models: TIGER-Lab/SWE-Next-7B, TIGER-Lab/SWE-Next-14B
• Project code: github.com/TIGER-AI-Lab/SWE-Next

Citation

@misc{liang2026swenextscalablerealworldsoftware,
      title={SWE-Next: Scalable Real-World Software Engineering Tasks for Agents},
      author={Jiarong Liang and Zhiheng Lyu and Zijie Liu and Xiangchao Chen and Ping Nie and Kai Zou and Wenhu Chen},
      year={2026},
      eprint={2603.20691},
      archivePrefix={arXiv},
      primaryClass={cs.SE},
      url={https://arxiv.org/abs/2603.20691},
}