SWE-Dev
收藏资源简介:
SWE-Dev是一个大规模数据集,旨在评估和训练自主编码系统在现实世界的功能开发任务上的能力。该数据集包含14,000个训练样本和500个测试样本,每个样本都提供了一个可运行的执行环境和开发者编写的可执行单元测试。SWE-Dev不仅为监督微调提供了高质量的数据,还通过可执行的单元测试提供了准确的奖励信号,支持强化学习。该数据集涵盖了17个聊天机器人大型语言模型、10个推理模型和10个多智能体系统,揭示了功能驱动开发是当前人工智能的挑战前沿。SWE-Dev为模型改进提供了一个有效的平台,通过在训练集上微调,使得一个70亿参数的模型在困难子集上与GPT-4o相当,凸显了其高质量训练数据的价值。
SWE-Dev is a large-scale dataset designed to evaluate and train autonomous coding systems on real-world functional development tasks. This dataset contains 14,000 training samples and 500 test samples, each paired with a runnable execution environment and executable unit tests written by developers. SWE-Dev not only provides high-quality data for supervised fine-tuning, but also offers accurate reward signals via executable unit tests to support reinforcement learning. This dataset covers 17 chatbot large language models, 10 reasoning models, and 10 multi-agent systems, revealing that function-driven development is the cutting-edge challenge of current artificial intelligence. SWE-Dev provides an effective platform for model improvement: fine-tuning a 7-billion-parameter model on the training set enables it to achieve performance comparable to GPT-4o on the hard subset, highlighting the value of its high-quality training data.
SWE-Dev 数据集概述
📌 基本信息
- 数据集名称: SWE-Dev
- 相关论文: SWE-Dev: Evaluating and Training Autonomous Feature-Driven Software Development
- 项目页面: Project Page
📚 数据集用途
- 用于评估和训练自主驱动的软件开发功能。
🛠 环境配置
- Python版本: 3.8或更高版本
- 操作系统支持: Windows和macOS
- 依赖安装: bash conda create -n swe_bench python=3.8 conda activate swe_bench git clone https://github.com/yifan-zhou1/SWE-Dev.git cd SWE-Dev pip install -r requirement.txt
🚦 使用步骤
-
运行
step1.py: bash python step1.py -
运行
step2.py: bash python step2.py -
运行
step3.py: bash python step3.py
📝 引用
- 使用该数据集时,请引用相关论文。




