microsoft/CoSAlign-Train
收藏Hugging Face2025-08-01 更新2025-05-31 收录
下载链接：
https://hf-mirror.com/datasets/microsoft/CoSAlign-Train
下载链接
链接失效反馈官方服务：
资源简介：
CoSAlign-Train是一个大规模合成的偏好训练数据集，用于训练大型语言模型（LLMs）进行可控的安全对齐（CoSA），以实现在推理时针对多样化的安全需求的细粒度适应。该数据集包括122K个合成的偏好对（配置、提示、选择的响应和拒绝的响应），来源于BeaverTails和WildGuard数据集，并跨越8种安全风险类型。每个响应对都根据安全合规性的帮助性进行评分。
CoSAlign-Train is a large-scale synthetic preference training dataset designed for training large language models (LLMs) for controllable safety alignment (CoSA), enabling fine-grained inference-time adaptation to diverse safety requirements. The dataset includes 122K synthetic preference pairs (config, prompt, chosen response, and rejected response) sourced from the BeaverTails and WildGuard datasets, spanning across 8 safety risk types. Each response pair is scored based on safety-compliant helpfulness.
提供机构：
microsoft
