AI45Research/SATraj-OS

SATraj-OS是一个用于计算机使用代理（CUAs）的大规模GUI轨迹数据集，旨在支持能力学习和安全训练。该数据集包含真实的GUI截图、多步交互轨迹和对抗性安全场景。数据集分为能力轨迹和安全轨迹两类，能力轨迹捕捉任务的成功与失败，安全轨迹则记录代理是否检测并避免攻击或未能防御攻击。数据集包含10,496个轨迹，158,196个步骤，平均每个轨迹长度为15.07步。适用于多模态下一步预测、长程GUI规划、联合能力与安全训练等应用。

SATraj-OS is a large-scale GUI trajectory dataset for Computer-Using Agents (CUAs), designed for both capability learning and safety training. The dataset includes real GUI screenshots, multi-step interaction trajectories, and adversarial safety scenarios. It contains two types of trajectories: Capability and Safety. Capability trajectories capture task success and failure, while Safety trajectories capture whether an agent detects and avoids attacks or fails to defend against them. The dataset consists of 10,496 trajectories, 158,196 steps, with an average trajectory length of 15.07 steps. It is designed for applications such as multimodal next-step prediction, long-horizon GUI planning, and joint capability and safety training.