MYPCBENCH
收藏资源简介:
MYPCBENCH是由卡内基梅隆大学团队创建的一个用于评估个性化智能计算机使用代理的基准数据集,其核心在于构建了一个模拟真实个人计算机环境的可复现Linux桌面系统。该数据集包含184个任务,覆盖了17个预登录的模拟网络应用及完整桌面堆栈,数据总量包括1,812笔银行交易、2,398封电子邮件等丰富记录,所有数据均基于电视剧《办公室》中的角色Michael Scott的个人身份生成,确保跨应用一致性。数据集的创建过程通过确定性生成管道,从统一的用户身份规范中同步生成多应用数据,并利用编码代理进行质量验证。该数据集主要应用于评估人工智能代理在真实个人计算机环境中的个性化任务执行能力,旨在解决现有基准测试在模拟用户历史数据和跨应用上下文方面的不足,推动个性化助手技术的发展。
MYPCBENCH is a benchmark dataset developed by the team at Carnegie Mellon University for evaluating personalized intelligent computer usage agents. Its core component is a reproducible Linux desktop system that simulates a real-world personal computing environment. This dataset includes 184 tasks, covering 17 pre-logged simulated web applications and a complete desktop stack. It contains a rich set of records, such as 1,812 bank transactions and 2,398 emails, among other data. All data is generated based on the personal identity of Michael Scott, a character from the TV series *The Office*, to ensure cross-application consistency. The dataset is created via a deterministic generation pipeline, which synchronously generates multi-application data from a unified user identity specification, and utilizes coding agents for quality validation. This dataset is primarily used to assess the personalized task execution capabilities of AI agents within real personal computer environments, aiming to address the limitations of existing benchmarks in simulating user historical data and cross-application context, thereby advancing the development of personalized assistant technologies.

- 1MyPCBench: A Benchmark for Personally Intelligent Computer-Use Agents卡内基梅隆大学 · 2026年



