leosltl/Android-in-the-Zoo-mirror

AITZ (Android-in-the-Zoo) 是一个基于AITW (Android-in-the-Wild)的GUI-agent基准测试数据集。它提供了18,643个屏幕-动作对，覆盖约2,500条指令，每条指令都标注了Chain-of-Action-Thought (CoAT)语义标签，包括屏幕描述、动作思考、动作描述和结果屏幕描述。数据集的结构包括训练集和测试集，分别包含不同的子集。训练集包括general、google_apps、install、single和web_shopping子集，其中single子集仅用于训练。测试集包括general、google_apps、install和web_shopping子集，与AITW的保留子集相匹配。每个episode目录包含步骤截图和带有CoAT注释的JSON文件。

AITZ is a GUI-agent benchmark built on top of AITW (Android-in-the-Wild). It provides 18,643 screen–action pairs over ~2,500 instructions, each annotated with Chain-of-Action-Thought (CoAT) semantic labels: screen description, action thinking, action description, and the resulting screen description. The dataset structure includes train and test splits with different subsets. The train split includes general, google_apps, install, single (train-only), and web_shopping subsets. The test split includes general, google_apps, install, and web_shopping subsets, matching AITWs held-out subsets. Each episode directory contains step screenshots and JSON files with CoAT annotations.