synthux-economy-r5-w062
收藏资源简介:
SynthUX Computer-Use数据集(economy sim r5版本)是一个用于计算机使用和图形用户界面(GUI)代理任务的合成数据集。该数据集旨在通过模拟真实公司环境中的员工设备会话来生成基于视觉的计算机使用轨迹。每个数据记录代表一名员工在模拟公司内的完整会话过程。其核心机制是将自然语言目标扩展为节点树,其中终端节点通过低级别的鼠标和键盘输入,驱动一系列真实的桌面模拟器应用程序(包括终端、笔记、VS Code、浏览器、Slack/Teams、邮件、表格、幻灯片等)。整个交互过程被完整捕获,形成两种主要数据形式:全过程的屏幕录制视频,以及每个终端节点在最后输入事件时刻的静态截图帧。数据集采用应用原生的多应用程序执行模式,无需工作台覆盖层,并强调每个参与者(actor)的内容多样性、跨操作系统(macOS、Windows、浏览器OS)的消息传递,以及大型多公司经济模拟的构建。每个经济模拟会作为独立的数据集仓库发布。数据文件包括按行存储轨迹记录的JSONL文件、按轨迹和节点ID组织的PNG截图文件夹,以及WebM格式的屏幕录像文件夹。记录模式详细定义了轨迹ID、参与者与组织信息(ID、公司、角色、职位)、模拟环境、自然语言目标、节点树结构、密集的低级输入事件序列、因果模拟器状态观测、观测与事件的节点对齐信息,以及媒体文件引用。本构建版本包含36条轨迹、36个视频、834帧图像,涉及6个不同的模拟公司,环境分布为浏览器OS(10条)、macOS网页模拟(8条)和Windows网页模拟(18条)。数据由SynthUX系统生成,确保观测的因果性,仅使用低级输入作为突变通道,未在捕获阶段引入高级应用突变。该数据集适用于训练和评估能够理解并执行复杂、多步骤计算机任务的GUI智能体。
The SynthUX Computer-Use dataset (economy sim r5 version) is a synthetic dataset for computer use and graphical user interface (GUI) agent tasks. It aims to generate vision-based computer use trajectories by simulating employee device sessions in realistic company environments. Each data record represents a complete session of an employee within a simulated company. Its core mechanism expands natural language goals into a node tree, where terminal nodes drive a series of real desktop simulator applications (including terminal, notes, VS Code, browser, Slack/Teams, email, spreadsheets, slides, etc.) through low-level mouse and keyboard inputs. The entire interaction process is fully captured, forming two main data formats: screen recording videos of the entire process, and static screenshot frames at the moment of the final input event for each terminal node. The dataset employs an application-native multi-application execution mode without workspace overlays, emphasizing content diversity per actor, cross-operating system (macOS, Windows, browser OS) messaging, and the construction of large multi-company economy simulations. Each economy simulation is released as an independent dataset repository. Data files include JSONL files storing trajectory records line-by-line, PNG screenshot folders organized by trajectory and node IDs, and WebM format screen recording folders. The recording mode details trajectory ID, participant and organizational information (ID, company, role, position), simulation environment, natural language goal, node tree structure, dense low-level input event sequences, causal simulator state observations, node alignment information for observations and events, and media file references. This build version contains 36 trajectories, 36 videos, 834 frames, involving 6 different simulated companies, with environment distribution as browser OS (10), macOS web simulation (8), and Windows web simulation (18). The data is generated by the SynthUX system, ensuring causal observations, using only low-level inputs as mutation channels without introducing high-level application mutations during capture. This dataset is suitable for training and evaluating GUI agents capable of understanding and executing complex, multi-step computer tasks.
SynthUX Computer-Use 数据集 — economy sim r5
数据集概述
这是一个基于 SynthUX 生成的接地视觉计算机使用轨迹数据集,模拟了一家虚拟公司内员工的设备会话。每条记录代表一个员工的桌面会话,包含从目标到操作的完整节点树,最终通过低层级鼠标/键盘操作驱动真实桌面模拟器应用,并记录完整的屏幕视频和逐节点图像帧。
文件结构
data.jsonl— 每条渲染轨迹的记录screenshots/<trajectory>/<node_id>.png— 每个终端节点在最后一次输入事件时的图像帧videos/<trajectory>.webm— 完整的屏幕录制视频
记录字段说明
| 字段 | 说明 |
|---|---|
trajectory_id |
唯一标识符(<actor>__<env>) |
actor_id, company, role_id, title |
员工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 |
员工的终端节点(包含 surface, action, target, args) |
trajectory.input_events |
密集的低层级鼠标/键盘/脚本事件,含时间戳 t_ms |
trajectory.observations |
每个节点后的因果模拟器状态 |
trajectory.alignment |
将观测和输入事件范围与节点 ID 关联 |
media.video |
屏幕录制视频(webm 格式) |
media.frames |
逐节点 PNG 图像帧(含节点 ID、时间戳、二进制引用) |
数据集统计(当前版本)
- 轨迹数量:36 条
- 视频数量:36 个
- 图像帧数量:834 帧
- 公司数量:6 家
- 按模拟器环境分布:
- browser-os:10 条
- macos-web-next:8 条
- windows-web-next:18 条
数据特点
- 原生多应用执行(无工作台叠加层)
- 每个参与者的内容多样性
- 跨操作系统消息传递
- 大型多公司经济体仿真
许可证
Apache-2.0




