Recraft-V2
收藏arXiv2025-09-30 收录
下载链接：
https://huggingface.co/datasets/Rapidata/Recraft-V2_t2i_human_preference
下载链接
链接失效反馈官方服务：
资源简介：
该数据集名为Recraft-V2，包含了13,000个人工标注的图像-文本对，这些对涵盖了三个关键维度：对齐度、连贯性和偏好，提供了多维度标注。数据的多样性确保了奖励学习的健壮性，为衡量对抗性扰动效应提供了一个可靠的基准。该数据集的规模为13,000对，任务是对奖励模型进行预训练。
This dataset, named Recraft-V2, contains 13,000 manually annotated image-text pairs with multi-dimensional annotations spanning three core dimensions: alignment, coherence, and preference. The diversity of the data ensures the robustness of reward learning and provides a reliable benchmark for assessing the effects of adversarial perturbations. With a total of 13,000 pairs, this dataset is intended for the pre-training of reward models.
