jacob-valdez/synthux-economy-r5-w023
收藏资源简介:
SynthUX计算机使用数据集(经济模拟r5版)是一个基于视觉的计算机使用轨迹数据集,由SynthUX生成。每个记录模拟了一个公司内一个工作者的设备会话:目标扩展为节点树,终端节点通过低级别鼠标/键盘输入驱动真实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),观察到的轨迹被记录为屏幕视频和每个节点的帧。该数据集支持应用原生多应用程序执行(无工作台覆盖),具有每个参与者的内容多样性、跨操作系统消息传递和大型多公司经济模拟功能。每个经济模拟都作为独立的数据集仓库发布。数据内容包括JSONL格式的记录文件、每个轨迹的截图(按节点存储为PNG格式)和完整屏幕录制视频(WebM格式)。记录模式涵盖轨迹ID、参与者信息、环境、自然语言目标、节点树、输入事件、观察结果、对齐和媒体文件。统计信息显示本版本包含36个轨迹、955个帧、6家公司,环境分布为browser-os(8个)、macos-web-next(19个)和windows-web-next(9个)。数据由SynthUX生成,观察结果是因果的(突变通道为低级别输入),在捕获过程中未使用高级别应用程序突变。
SynthUX Computer Usage Dataset (Economic Simulation r5 Version) is a vision-based computer interaction trajectory dataset generated by SynthUX. Each record simulates a device session of a worker within a company: the task goal is expanded into a node tree, and terminal nodes drive real desktop simulator applications (including terminals, note-taking tools, VS Code, web browsers, Slack/Teams, email clients, spreadsheets, presentation software, etc.) via low-level mouse and keyboard inputs. The observed interaction trajectories are recorded as screen videos and per-node frame sequences. This dataset supports native multi-application execution (without workspace overlay), and features content diversity per participant, cross-operating system messaging functionality, and large-scale multi-company economic simulation capabilities. Each economic simulation is released as an independent dataset repository. The dataset content includes JSONL-formatted record files, per-trajectory screenshots stored as PNG files organized by node, and full-screen recording videos in WebM format. The record schema covers trajectory ID, participant information, environment configuration, natural language task goals, node trees, input events, observation data, alignment annotations, and media files. Statistical summaries show that this version contains 36 trajectories, 955 frames, across 6 companies, with the environment distribution being browser-os (8 instances), macos-web-next (19 instances), and windows-web-next (9 instances). The dataset is generated by SynthUX, where the observational data is causal (the mutation channel utilizes low-level inputs), and no high-level application-level mutations were used during the data capture process.




