synthux-economy-r5-w054
收藏资源简介:
SynthUX Computer-Use Dataset 是一个基于视觉的计算机使用轨迹合成数据集,专注于经济模拟场景(economy sim r5)。该数据集由 SynthUX 生成,旨在模拟真实工作环境中用户与多应用程序桌面交互的轨迹。每个数据记录代表一个虚拟工作者在模拟公司内的设备会话:自然语言目标被扩展为节点树,终端节点通过低级别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、VS Code、浏览器、Slack/Teams、邮件、表格、幻灯片等)。交互轨迹被记录为屏幕视频以及每个终端节点在最后输入事件时刻的截图帧。数据集包含三个部分:data.jsonl 文件(每条轨迹对应一个 JSON Lines 记录)、screenshots/ 目录(存储每个轨迹中每个终端节点的 PNG 格式截图)和 videos/ 目录(存储完整的屏幕录制视频,WebM 格式)。记录模式详细定义了多个字段,如轨迹唯一标识、参与者信息、模拟器环境、自然语言目标、节点树结构、低级别输入事件序列、因果观察状态、对齐信息和媒体引用。数据规模为36条轨迹,对应36个视频和1090个截图帧,涉及6个不同的模拟公司,并按模拟环境分布(browser-os 12条、macos-web-next 15条、windows-web-next 9条)。数据集采用应用原生多应用程序执行模式,无工作台覆盖,支持跨操作系统消息传递,并模拟大型多公司经济环境。该数据集适用于图形用户界面(GUI)智能体训练、计算机使用行为分析、多任务自动化以及人机交互研究等任务,特别适合需要低级别输入事件和屏幕视觉观察的强化学习或模仿学习场景。数据生成过程中仅使用低级别输入作为突变通道,未依赖高级别应用程序突变,确保了观察的因果性。
The SynthUX Computer-Use Dataset is a vision-based synthetic dataset of computer-use trajectories, focusing on economy simulation scenarios (economy sim r5). It is generated by SynthUX to simulate user interactions with multi-application desktops in real work environments. Each data record represents a device session of a virtual worker within a simulated company: a natural language goal is expanded into a node tree, with terminal nodes driving real desktop simulator applications (including terminal, notes, VS Code, browser, Slack/Teams, email, spreadsheets, slides, etc.) via low-level mouse and keyboard inputs. The observed interaction trajectories are recorded as screen videos and screenshot frames for each terminal node at the moment of the final input event. The dataset consists of three parts: 1) a data.jsonl file with one JSON Lines record per trajectory, 2) a screenshots/ directory storing PNG-format screenshots for each terminal node in each trajectory, and 3) a videos/ directory storing complete screen recording videos in WebM format. The recording schema defines multiple fields, including a unique trajectory identifier, participant information (e.g., company, role, position), simulator environment (macos-web-next, windows-web-next, or browser-os), natural language goal, node tree structure, low-level input event sequences, causal observation states, alignment information, and media references. This version includes 36 trajectories, corresponding to 36 videos and 1090 screenshot frames, involving 6 different simulated companies, and distributed across simulation environments (browser-os 12, macos-web-next 15, windows-web-next 9). The dataset employs an application-native multi-application execution mode without workbench overlays, supports cross-OS messaging, and simulates a large-scale multi-company economy. Each independent economy simulation is released as a separate dataset repository. It is suitable for tasks such as graphical user interface (GUI) agent training, computer-use behavior analysis, multi-task automation, and human-computer interaction research, particularly for reinforcement learning or imitation learning scenarios requiring low-level input events and screen visual observations. The data generation process uses only low-level inputs as mutation channels, avoiding high-level application mutations to ensure causal observability.
数据集概述
数据集名称:SynthUX Computer-Use (economy sim) r5
许可证:Apache-2.0
任务类别:其他
标签:计算机使用、GUI代理、合成数据、多应用
数据集描述
该数据集包含由 SynthUX 生成的基于视觉的计算机使用轨迹。每条记录对应模拟公司中一名员工的设备会话:一个目标展开为节点树,终端节点通过底层鼠标/键盘输入驱动真实桌面模拟器应用(如终端、笔记、VS Code、浏览器、Slack/Teams、邮件、表格、幻灯片等),并记录观察到的轨迹,包括屏幕视频和每个节点的帧。
特点:
- 应用原生多应用执行(无工作台覆盖)
- 每个角色的内容多样性
- 跨操作系统的消息传递
- 大型多公司经济体
数据集内容
data.jsonl:每条渲染轨迹的记录。screenshots/<trajectory>/<node_id>.png:每个终端节点在其最后一次输入事件时的帧。videos/<trajectory>.webm:完整的屏幕录制。
记录模式
| 字段 | 含义 |
|---|---|
trajectory_id |
唯一 ID(<actor>__<env>) |
actor_id, company, role_id, title |
谁,在哪个组织中 |
env |
模拟器:macos-web-next / windows-web-next / browser-os |
goal |
会话的自然语言目标 |
tree.nodes |
角色的终端节点(surface, action, target, args) |
trajectory.input_events |
密集的底层鼠标/键盘/脚本事件,附带 t_ms |
trajectory.observations |
每个节点后的因果模拟器状态 |
trajectory.alignment |
将观察结果和输入事件范围链接到节点 ID |
media.video |
屏幕录制(webm) |
media.frames |
每节点的 PNG 帧(node_id, t_ms, blob_ref) |
统计信息(当前构建)
- 轨迹数量:36
- 视频数量:36 · 帧数量:1090
- 公司数量:6
- 按环境分布:browser-os(12),macos-web-next(15),windows-web-next(9)
来源
使用 SynthUX(synthux.multicompany.day.v1)生成。观察结果是因果性的(突变通道 = low_level_input);捕获过程中未使用高级应用突变。




