Safer-Instruct (SI)
收藏arXiv2025-09-30 收录
下载链接：
https://github.com/uscnlp-lime/safer-instruct
下载链接
链接失效反馈官方服务：
资源简介：
该数据集是为了构建语言模型的安全偏好而设计的，它包含了来自不同安全相关内容类别的数据。这些类别包括仇恨言论、自残内容、性内容以及非法活动等，数据来源于Reddit和Twitter等多样化平台。该数据集共有10,255个样本，横跨四个不同的类别，其任务是生成针对语言模型的安全偏好数据。
This dataset is designed for constructing safety preference data for language models. It contains data from multiple safety-related content categories, including hate speech, self-harm content, sexual content, illegal activities and more. The data is collected from diverse platforms such as Reddit and Twitter. In total, this dataset has 10,255 samples spanning four distinct categories, and its core objective is to generate safety preference data for language models.
提供机构：
Authors of the paper
