MolmoWeb-SyntheticSkills

MolmoWeb-SyntheticSkills 是一个合成网页导航技能的数据集，其中每个技能代表一个低级任务的轨迹（例如查找并打开、填写表单）。每个样本将一个指令与一系列网页截图及相应的代理操作（如点击、输入、滚动等）配对。数据集包含 5535 个训练样本和 10 个预览样本。主要特征包括：sample_id（唯一标识符）、instruction（JSON 编码的任务指令）、trajectory（JSON 编码的轨迹，包含代理操作和截图文件名）、images（截图原始数据列表）和 image_paths（截图路径列表）。轨迹的每个步骤包含截图文件名、代理操作（包括可解析的操作字符串、自然语言描述和结构化输出）、浏览器状态（当前 URL、页面索引等）以及操作时间戳。数据集适用于网页导航、代理操作模拟等研究任务，采用 ODC-BY 1.0 许可，仅供研究和教育使用。

MolmoWeb-SyntheticSkills is a dataset focused on synthetic web navigation skills, where each skill represents a trajectory of a low-level task, such as "find and open" or "fill out a form". Each sample pairs an instruction with a sequence of web page screenshots and corresponding agent actions (e.g., click, input, scroll, etc.). The dataset contains 5535 training samples and 10 preview samples. Its core features include: sample_id (unique identifier), instruction (JSON-encoded task instruction), trajectory (JSON-encoded trajectory containing agent actions and screenshot filenames), images (list of raw screenshot data), and image_paths (list of screenshot file paths). Each step in the trajectory includes the screenshot filename, agent actions (covering parsable action strings, natural language descriptions, and structured outputs), browser state (such as current URL, page index, etc.), and operation timestamp. This dataset is applicable to research tasks including web navigation and agent action simulation, and is licensed under ODC-BY 1.0 for research and educational use only.