GAIR/daVinci-Agency

daVinci-Agency是一个高质量的数据集，专为训练代理执行长期软件工程任务而设计。该数据集通过从真实世界的Pull Request（PR）链中挖掘结构化监督，提供了明确建模软件演化过程的轨迹。数据集通过三个相互关联的机制实现这一目标：渐进式任务分解（将复杂目标转化为可验证的提交单元）、长期一致性执行（通过统一目标在100多个工具调用中保持功能一致性）和可验证的细化（从真实世界的错误修复轨迹中学习真实的迭代模式）。数据集采用多轮交互格式，包含系统提示、用户查询和工具定义，适用于代理微调。尽管样本量较小（239个样本），但在多个基准测试中表现优异，优于更大的合成数据集。

daVinci-Agency is a high-quality dataset designed for training agents on long-horizon software engineering tasks. By mining structured supervision from real-world Pull Request (PR) chains, the dataset provides trajectories that explicitly model the software evolution process through three interlocking mechanisms: Progressive Task Decomposition (transforming complex objectives into verifiable submission units), Long-term Consistency Enforcement (maintaining functional coherence through unified objectives across 100+ tool invocations), and Verifiable Refinement (learning authentic iterative patterns from real-world bug-fix trajectories). The dataset is formatted for Agentic Fine-tuning, representing each evolutionary trajectory as a multi-turn interaction with system prompts, user queries, and tool definitions. Despite its small sample size (239 samples), it consistently outperforms much larger synthetic datasets in various benchmarks.