hny-research/hny-vcv-ui
收藏资源简介:
HNY-VCV-UI是一个合成用户界面(UI)视觉合同验证数据集,专门用于视觉问答任务。该数据集包含训练集和验证集,用于测试紧凑视觉观察者模型是否能根据提供的图像和视觉合同判断合同是否被满足(SATISFIED)、违反(VIOLATED)或未知(UNKNOWN),并提供基于证据的决策。数据集由HNY研究部门通过实验002的合成UI视觉合同管道生成,包括132个训练样本和44个验证样本,总计176个PNG图像文件,每个样本包含id、图像路径、元数据(如决策、难度、分割和图像哈希)、提示和目标。
HNY-VCV-UI is a synthetic UI visual-contract verifier SFT dataset for visual-question-answering tasks. It is used to test whether a compact visual observer can determine if an image satisfies, violates, or has an unknown status relative to a visual contract, and provide grounded evidence. The dataset is generated by the HNY Research Division Experiment 002 synthetic UI visual-contract pipeline, consisting of 132 training rows and 44 validation rows, with a total of 176 PNG image files. Each JSONL row includes id, image_path, metadata (such as decision, difficulty, split, and image hash), prompt, and target.




