ProGuard Dataset
收藏资源简介:
ProGuard数据集是由上海人工智能实验室、南京大学PRLab和北京航空航天大学联合构建的多模态安全标注数据集,包含87,000条经过严格标注的文本、图像及图文混合样本。该数据集采用分层多模态安全分类体系,每条数据均标注有二元安全标签和细粒度风险类别,有效缓解了传统方法中的模态偏差问题。数据来源整合了GuardReasoner、BeaverTails-V等10个权威安全数据集,通过大模型多数投票机制进行标注验证,人类验证准确率超过90%。该数据集专为训练主动式安全防护模型设计,可应用于生成式AI的内容安全过滤、风险分类和未知威胁检测等领域。
The ProGuard Dataset is a multimodal safety annotation dataset jointly constructed by the Shanghai AI Laboratory, Nanjing University PRLab, and Beihang University. It contains 87,000 rigorously annotated text, image, and text-image hybrid samples. Adopting a hierarchical multimodal safety classification framework, each sample in this dataset is annotated with binary safety labels and fine-grained risk categories, effectively mitigating the modal bias problem prevalent in traditional approaches. The dataset compiles 10 authoritative safety datasets including GuardReasoner and BeaverTails-V, with its annotation validation conducted via the majority voting mechanism of large language models (LLMs), achieving a human verification accuracy rate of over 90%. This dataset is specifically designed for training proactive safety defense models, and can be deployed in scenarios such as content safety filtering, risk classification, and unknown threat detection for generative AI.

- 1ProGuard: Towards Proactive Multimodal Safeguard上海人工智能实验室; 南京大学·PRLab; 北京航空航天大学 · 2025年



