SAFESORA
收藏资源简介:
SAFESORA数据集是首个捕捉文本到视频生成任务中真实人类偏好的数据集,由北京大学人工智能安全与治理研究中心创建。该数据集包含14,711个独特的文本提示和57,333个由四个不同的大型视觉模型生成的独特视频。数据集的创建过程涉及两阶段注释,引导众包工作者根据自身感知解释帮助性和无害性概念。SAFESORA数据集的应用领域广泛,包括训练文本-视频审查模型和通过微调提示增强模块或扩散模型来调整大型视觉模型,旨在解决模型输出与人类价值观对齐的问题。
The SAFESORA dataset, developed by the AI Security and Governance Research Center of Peking University, is the first dataset that captures real human preferences in text-to-video generation tasks. This dataset contains 14,711 unique text prompts and 57,333 distinct videos generated by four different large-scale vision models. The dataset construction involves a two-stage annotation process, where crowdworkers are guided to interpret the concepts of helpfulness and harmlessness based on their own perceptions. The SAFESORA dataset has a wide range of application scenarios, including training text-video moderation models and aligning large-scale vision models via fine-tuning prompt enhancement modules or diffusion models, aiming to address the problem of aligning model outputs with human values.




