synthux-economy-r5-w066
收藏资源简介:
SynthUX Computer-Use数据集(economy sim r5版本)是一个合成的、多应用程序的计算机使用轨迹数据集,专门设计用于图形用户界面(GUI)代理和计算机使用研究。该数据集模拟了公司环境中工作人员的设备会话,通过SynthUX系统生成基于视觉的交互轨迹。每个记录代表一个完整的用户会话,其中自然语言目标被扩展为节点树,终端节点通过低级鼠标和键盘输入驱动真实的桌面模拟器应用程序,包括终端、笔记软件、VS Code、浏览器、Slack/Teams、邮件、表格、幻灯片等。数据集采用应用程序原生的多应用程序执行模式,无需工作台覆盖,确保交互的真实性。数据集包含36个轨迹,涉及6个不同的虚拟公司,覆盖三种模拟环境:browser-os(10个轨迹)、macos-web-next(12个轨迹)和windows-web-next(14个轨迹)。每个轨迹记录包括完整的屏幕录制视频(WebM格式)、每个终端节点在最后输入事件时刻捕获的PNG帧图像,以及结构化的JSONL元数据。元数据字段包括轨迹ID、参与者信息(演员ID、公司、角色ID、职位)、模拟环境、自然语言目标、节点树结构(包含表面、动作、目标、参数)、密集的低级输入事件序列(带时间戳)、因果模拟器状态观察结果,以及观察结果与节点ID的对应关系。数据集旨在支持跨操作系统消息传递、多公司经济模拟的研究,适用于GUI自动化、人机交互分析、任务导向对话系统训练等场景。数据生成过程中仅使用低级输入突变,确保观察结果的因果性,未采用高级应用程序突变,提高了数据的可靠性和可复现性。
SynthUX Computer-Use Dataset (economy sim r5 version) is a synthetic, multi-application computer interaction trajectory dataset specifically designed for graphical user interface (GUI) agents and computer use research. This dataset simulates device sessions of workers in corporate environments, generating vision-based interaction trajectories via the SynthUX system. Each record represents a complete user session, where natural language goals are expanded into node trees, and terminal nodes drive real desktop emulator applications via low-level mouse and keyboard inputs, including terminals, note-taking software, VS Code, browsers, Slack/Teams, email, spreadsheets, presentations, etc. The dataset adopts the native multi-application execution mode of applications without desktop overlay, ensuring the authenticity of interactions. The dataset contains 36 trajectories across 6 distinct virtual companies, covering three simulation environments: browser-os (10 trajectories), macos-web-next (12 trajectories), and windows-web-next (14 trajectories). Each trajectory record includes full screen recording videos (in WebM format), PNG frame images captured at the final input event moment for each terminal node, and structured JSONL metadata. The metadata fields include trajectory ID, participant information (actor ID, company, role ID, job title), simulation environment, natural language goals, node tree structure (including surface, action, target, parameters), dense low-level input event sequences with timestamps, causal simulator state observations, and the correspondence between observations and node IDs. This dataset is intended to support research on cross-operating system messaging and multi-company economic simulation, and is applicable to scenarios such as GUI automation, human-computer interaction analysis, and task-oriented dialogue system training. Only low-level input mutations are used during the data generation process to ensure the causality of observations, and high-level application mutations are not adopted, improving the reliability and reproducibility of the dataset.
数据集概述
SynthUX Computer-Use Dataset — economy sim r5 是一个由 SynthUX 生成的、基于视觉的计算机操作轨迹数据集,模拟了经济模拟环境中的多应用桌面操作。
数据集内容
data.jsonl:每条记录对应一个渲染后的操作轨迹。screenshots/<trajectory>/<node_id>.png:每个终端节点在其最后一次输入事件时的截图。videos/<trajectory>.webm:完整的屏幕录制视频。
记录字段说明
| 字段 | 含义 |
|---|---|
trajectory_id |
唯一标识(<actor>__<env>) |
actor_id, company, role_id, title |
操作者身份及所属组织信息 |
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 |
操作者的终端节点(包含 surface, action, target, args) |
trajectory.input_events |
密集的低级鼠标/键盘/脚本事件及时间戳 t_ms |
trajectory.observations |
每个节点后的因果模拟器状态 |
trajectory.alignment |
将观测与输入事件范围关联到节点ID |
media.video |
屏幕录制视频(webm格式) |
media.frames |
每个节点的PNG截图(包含节点ID、时间戳、blob引用) |
数据集规模(本次构建)
- 轨迹数量:36
- 视频数量:36,截图数量:1049
- 模拟公司数量:6
- 各模拟器轨迹分布:
browser-os10条,macos-web-next12条,windows-web-next14条
数据来源与生成
- 使用 SynthUX 生成(版本:synthux.multicompany.day.v1)。
- 观测过程是因果性的(突变通道为低级输入
low_level_input),捕获期间未使用高级应用突变。 - 数据集遵循 Apache-2.0 许可证。




