synthux-economy-r5-w055
收藏资源简介:
SynthUX Computer-Use Dataset(economy sim r5版本)是一个用于计算机使用与图形用户界面(GUI)智能体研究的合成数据集。该数据集通过SynthUX平台生成,模拟了多公司经济环境中的员工设备使用轨迹。其核心机制是将自然语言描述的工作目标分解为节点树,终端节点通过模拟低级别的鼠标和键盘输入,驱动一系列真实的桌面应用程序(包括终端、笔记、VS Code、浏览器、Slack/Teams、邮件、表格、幻灯片等),从而产生逼真的计算机交互序列。数据集旨在支持对多应用、跨平台、基于目标的GUI操作行为的研究。数据集内容包含三个部分:1) `data.jsonl`文件,每条JSON Lines记录对应一个完整的渲染轨迹;2) `screenshots`目录,为每个轨迹的每个终端节点提供在节点最后一次输入事件时捕获的PNG格式屏幕截图;3) `videos`目录,提供每个轨迹完整的屏幕操作录制的WebM格式视频。每条数据记录包含丰富的结构化信息:唯一的轨迹标识符;参与者信息(ID、所属公司、角色、职位);模拟环境(macos-web-next、windows-web-next或browser-os);自然语言描述的任务目标;完整的节点树结构(包含界面元素、动作、目标、参数);密集的时间戳化低级别输入事件序列;模拟器在每个节点执行后的因果状态观察;用于对齐观察、输入事件与节点的映射信息;以及指向视频和节点截图的媒体引用。本次发布的构建版本包含36个独立轨迹,对应36段屏幕视频和958张节点截图,涉及6家不同的模拟公司。轨迹分布在三种模拟环境中:browser-os(7个)、macos-web-next(13个)、windows-web-next(16个)。数据生成过程强调因果性和真实性,仅通过低级别输入通道驱动模拟器,未在捕获阶段使用高级应用内突变。
The SynthUX Computer-Use Dataset (economy sim r5 version) is a synthetic dataset for computer usage and graphical user interface (GUI) agent research. Generated via the SynthUX platform, it simulates employee device usage trajectories in a multi-company economic environment. Its core mechanism decomposes natural language-described work goals into node trees, where terminal nodes drive a series of real desktop applications (including terminal, notes, VS Code, browser, Slack/Teams, email, spreadsheets, slides, etc.) by simulating low-level mouse and keyboard inputs, thereby producing realistic computer interaction sequences. The dataset aims to support research on multi-application, cross-platform, goal-based GUI operation behaviors. The dataset content consists of three parts: 1) the `data.jsonl` file, where each JSON Lines record corresponds to a complete rendering trajectory; 2) the `screenshots` directory, providing PNG format screenshots captured at the last input event for each terminal node of each trajectory; 3) the `videos` directory, offering WebM format videos of complete screen operations for each trajectory. Each data record contains rich structured information: a unique trajectory identifier; participant information (ID, company, role, position); simulation environment (macos-web-next, windows-web-next, or browser-os); natural language-described task goals; a complete node tree structure (including interface elements, actions, goals, parameters); a dense timestamped low-level input event sequence; causal state observations from the simulator after each node execution; mapping information for aligning observations, input events, and nodes; and media references to videos and node screenshots. This release includes 36 independent trajectories, corresponding to 36 screen videos and 958 node screenshots, involving 6 different simulated companies. The trajectories are distributed across three simulation environments: browser-os (7), macos-web-next (13), and windows-web-next (16). The data generation process emphasizes causality and realism, driving the simulator solely through low-level input channels without using high-level in-application mutations during the capture phase.
数据集概况
SynthUX Computer-Use Dataset — economy sim r5 是一个模拟公司内部设备操作的计算机使用轨迹数据集。每个记录对应一个工作者的桌面会话,其目标被分解为节点树,终端节点通过低层级鼠标/键盘输入驱动真实桌面模拟器应用,最终生成屏幕视频和逐节点帧。
数据集许可
- Apache 2.0 许可证
数据内容
数据集包含以下文件:
- data.jsonl:每条记录对应一个渲染后的轨迹。
- screenshots/<trajectory>/<node_id>.png:每个终端节点在最后一次输入事件时的帧图像。
- videos/<trajectory>.webm:完整屏幕录像。
记录字段
| 字段 | 含义 |
|---|---|
trajectory_id |
唯一标识(格式:<actor>__<env>) |
actor_id, company, role_id, title |
所属人员、公司、角色、职位 |
env |
模拟器类型:macos-web-next / windows-web-next / browser-os |
goal |
会话的自然语言目标 |
tree.nodes |
终端节点信息(surface, action, target, args) |
trajectory.input_events |
密集的低层级鼠标/键盘/脚本事件及其时间戳(t_ms) |
trajectory.observations |
每个节点后的因果模拟器状态 |
trajectory.alignment |
将观测值与输入事件范围链接到节点ID |
media.video |
屏幕录像(webm格式) |
media.frames |
逐节点PNG帧(含节点ID、时间戳、blob引用) |
数据统计(本版本)
- 轨迹数量:36
- 视频数量:36 | 帧数量:958
- 公司数量:6
- 按环境分布:browser-os(7)、macos-web-next(13)、windows-web-next(16)
数据来源
使用 SynthUX(synthux.multicompany.day.v1)生成。观测是因果性的(突变通道为低层级输入),捕获过程中未使用高层级应用突变。




