TIGER-Lab/EditReward-Data

EditReward-数据集是一个大规模、高保真的指令引导图像编辑人类偏好数据集。该数据集包含超过20万对手动画注的偏好对，由经过培训的专家按照严格的标准化协议进行精心筛选，确保与考虑的人类判断高度一致并最小化标签噪声。数据集覆盖了由七种最先进模型生成的十二种不同来源的编辑，作为EditReward等评估指令引导图像编辑模型与人类偏好对齐情况的关键训练数据。EditReward-Data is a large-scale, high-fidelity human preference dataset for instruction-guided image editing. It contains over 200K manually annotated preference pairs, meticulously curated by trained experts following a rigorous and standardized protocol to ensure high alignment with human judgment and minimize label noise. The dataset covers a diverse range of edits produced by seven state-of-the-art models across twelve distinct sources, serving as crucial training data for reward models like EditReward designed to score instruction-guided image edits.