nevillathiya001/Orchard

Orchard数据集是一个开源代理建模框架的轨迹发布，包含两个并行代理建模数据集：swe子集和gui子集。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的最终补丁是否通过问题的隐藏测试套件（包括成功和失败的轨迹），用于文本生成、工具使用和代理任务。gui子集包含3,070个来自网页浏览GUI代理（pae-webvoyager基准）的成功每步展开，覆盖409个独特任务，每行是轨迹中的一个决策点，包含对话历史、截图和标量奖励，用于图像-文本到文本、工具使用和GUI代理任务。数据集遵循OpenAI风格的聊天模式，可变数据以JSON字符串存储在metadata列中，支持多模态和文本处理。The Orchard dataset is a trajectory release accompanying the open-source agentic modeling framework, comprising two parallel agentic-modeling datasets: the swe subset and the gui subset. The swe subset includes 107,185 multi-turn software-engineering trajectories across 2,788 GitHub repositories, each labeled with whether the agents final patch passed the issues hidden test suite (including both successful and unsuccessful trajectories), designed for text-generation, tool-use, and agent tasks. The gui subset includes 3,070 successful per-step rollouts from a web-browsing GUI agent (pae-webvoyager benchmark) across 409 unique tasks, where each row represents a decision point within a trajectory, containing conversation history up to that step, a screenshot, and a scalar reward, intended for image-text-to-text, tool-use, and GUI agent tasks. The dataset follows OpenAI-style chat schemas, with variable-shape side-data stored as JSON-encoded strings in a unified metadata column, supporting multimodal and text processing.