GUI-Libra
收藏资源简介:
GUI-Libra是由微软研究院与多所高校联合构建的图形用户界面推理数据集,包含8.1万条经过严格筛选的高质量样本。该数据集通过创新的构建与过滤流程,强化了推理轨迹与可执行动作的关联性,覆盖网页和移动端多平台交互场景。数据来源于现有开源GUI轨迹的增强与标注,重点解决动作对齐不足和噪声标签问题。其核心应用于提升原生GUI代理的决策能力,支持长视野任务导航中高层推理与底层执行的联合优化。
GUI-Libra is a graphical user interface (GUI) reasoning dataset jointly constructed by Microsoft Research and multiple universities, consisting of 81,000 high-quality samples rigorously curated. Through an innovative construction and filtering pipeline, this dataset strengthens the association between reasoning trajectories and executable actions, covering multi-platform interaction scenarios for both web and mobile interfaces. Derived from the enhancement and annotation of existing open-source GUI trajectories, the dataset primarily addresses the challenges of insufficient action alignment and noisy labels. Its core applications lie in enhancing the decision-making capabilities of native GUI agents and supporting joint optimization of high-level reasoning and low-level execution in long-horizon task navigation.



