microsoft/Orchard

Orchard数据集是一个伴随论文《Orchard: An Open-Source Agentic Modeling Framework》发布的轨迹数据集，包含两个从强大教师模型蒸馏出的并行代理建模数据集，均在Orchard Env沙盒基础设施中生成。具体包括：1. swe子集：包含107,185个多轮软件工程轨迹，覆盖2,788个GitHub仓库，每个轨迹标注了代理最终补丁是否通过隐藏测试套件（metadata.verify_status），包括成功和失败的轨迹；2. gui子集：包含3,070个来自Web浏览GUI代理（基于pae-webvoyager基准）的成功每步展开，覆盖409个独特任务，每个行代表轨迹中的一个决策点，包含到该步的对话、一张截图和经判断验证的标量奖励。两个子集均遵循OpenAI风格的聊天模式，可变形状的侧数据以JSON编码字符串形式存储在统一的metadata列中。The Orchard dataset is a trajectory dataset released alongside the paper *Orchard: An Open-Source Agentic Modeling Framework*. It contains two parallel agent modeling datasets distilled from powerful teacher models, both generated on the Orchard Env sandbox infrastructure. Specifically, it includes two subsets: 1. The SWE subset: It contains 107,185 multi-turn software engineering trajectories covering 2,788 GitHub repositories. Each trajectory is annotated with whether the agent's final patch passes the hidden test suite (metadata.verify_status), including both successful and failed trajectories; 2. The GUI subset: It contains 3,070 successful step-by-step rollouts from web-browsing GUI agents (based on the pae-webvoyager benchmark), covering 409 unique tasks. Each row represents a decision point in the trajectory, containing the dialogue up to that step, a screenshot, and a judgment-verified scalar reward. Both subsets follow the OpenAI-style chat format, and variable-shaped side data is stored as JSON-encoded strings in a unified metadata column.