AweAI-Team/DeNovoSWE-Trajectory-Filtered
收藏资源简介:
DeNovoSWE是一个用于从零开始生成整个代码仓库的数据集,旨在扩展长视野环境。它包含4,818个高质量实例,这些实例是从34,000个DeepSeek-v4-High轨迹中筛选出的11,000个轨迹。数据集提供了代码仓库重建所需的信息,包括实例ID、真实文档、PyPI包名、Docker镜像URL、GitHub用户和仓库名、工作目录路径、单元测试标识符列表、测试代码补丁以及测试二进制文件。该数据集在BeyondSWE-Doc2Repo任务上表现出色,训练自Qwen3.5-35B-A3B模型达到50%的性能。
DeNovoSWE is a dataset designed for generating entire code repositories from scratch, with the objective of expanding long-horizon environments. It comprises 4,818 high-quality instances, which are selected from 11,000 trajectories filtered from a total of 34,000 DeepSeek-v4-High trajectories. The dataset provides all necessary information for code repository reconstruction, including instance ID, ground-truth documentation, PyPI package names, Docker image URLs, GitHub usernames and repository names, working directory paths, lists of unit test identifiers, test code patches, and test binaries. It delivers excellent performance on the BeyondSWE-Doc2Repo task, achieving 50% performance when trained using the Qwen3.5-35B-A3B model.




