DeNovoSWE: Scaling Long-Horizon Environments for Generating Entire Repositories from Scratch
9
7 commits
2 linked in READMEs
updated Jun 14, 2026
DeNovoSWE is a large-scale dataset for whole-repository generation, as introduced in the paper "DeNovoSWE: Scaling Long-Horizon Environments for Generating Entire Repositories from Scratch".

| Field | Description |
|---|---|
instance_id | Unique identifier for each benchmark instance. |
document | Ground-truth documentation provided to the agent for repository reconstruction. |
pypi_name | PyPI package name used to enforce anti-cheating constraints during evaluation. |
image_url | URL of the Docker image configured for the environment. |
user | GitHub username or organization owning the repository. |
repo | Name of the target GitHub repository. |
workdir | Working directory path of the repository inside the Docker container. |
unit_test | List of all unit test identifiers. |
test_patch | Code patch for unit tests, applied during the evaluation phase. |
test_binary_files | Binary files used by unit tests that are unsuitable for standard text patching. |

Key parameters can be seen below:
| Parameter | Value |
|---|---|
| Max turns | 500 |
| Max sequence length | 256k |
| Temperature | 1 |
If you find this project useful for your research, please consider citing our paper:
@misc{zhao2026denovoswescalinglonghorizonenvironments,
title={DeNovoSWE: Scaling Long-Horizon Environments for Generating Entire Repositories from Scratch},
author={Jiale Zhao and Guoxin Chen and Fanzhe Meng and Wayne Xin Zhao and Ruihua Song and Ji-Rong Wen and Kai Jia},
year={2026},
eprint={2606.10728},
archivePrefix={arXiv},
primaryClass={cs.SE},
url={https://arxiv.org/abs/2606.10728},
}
This project is licensed under the CC BY 4.0 License - see the LICENSE file for details.
If you find this project useful, please consider giving it a β !
DeNovoSWE: Scaling Long-Horizon Environments for Generating Entire Repositories from Scratch
9
7 commits
2 linked in READMEs
updated Jun 14, 2026
DeNovoSWE is a large-scale dataset for whole-repository generation, as introduced in the paper "DeNovoSWE: Scaling Long-Horizon Environments for Generating Entire Repositories from Scratch".

| Field | Description |
|---|---|
instance_id | Unique identifier for each benchmark instance. |
document | Ground-truth documentation provided to the agent for repository reconstruction. |
pypi_name | PyPI package name used to enforce anti-cheating constraints during evaluation. |
image_url | URL of the Docker image configured for the environment. |
user | GitHub username or organization owning the repository. |
repo | Name of the target GitHub repository. |
workdir | Working directory path of the repository inside the Docker container. |
unit_test | List of all unit test identifiers. |
test_patch | Code patch for unit tests, applied during the evaluation phase. |
test_binary_files | Binary files used by unit tests that are unsuitable for standard text patching. |

Key parameters can be seen below:
| Parameter | Value |
|---|---|
| Max turns | 500 |
| Max sequence length | 256k |
| Temperature | 1 |
If you find this project useful for your research, please consider citing our paper:
@misc{zhao2026denovoswescalinglonghorizonenvironments,
title={DeNovoSWE: Scaling Long-Horizon Environments for Generating Entire Repositories from Scratch},
author={Jiale Zhao and Guoxin Chen and Fanzhe Meng and Wayne Xin Zhao and Ruihua Song and Ji-Rong Wen and Kai Jia},
year={2026},
eprint={2606.10728},
archivePrefix={arXiv},
primaryClass={cs.SE},
url={https://arxiv.org/abs/2606.10728},
}
This project is licensed under the CC BY 4.0 License - see the LICENSE file for details.
If you find this project useful, please consider giving it a β !