responsible-ai-labs/RAIL-HH-10K

RAIL-HH-10K 是一个多维度安全对齐数据集，包含 10,000 个高质量示例，覆盖 8 个伦理维度，其中 99.5% 的示例被标注。该数据集旨在促进 AI 安全和伦理研究，特别适合安全对齐研究、可解释 AI、基准测试和模型训练。数据集分为训练集、验证集和测试集，每个示例都包含详细的标注和评分，以及拒绝和选择的响应。RAIL-HH-10K 数据集在安全和用户影响方面取得了显著的改进，并提供了多维度安全评分工具，以帮助研究人员评估其数据。RAIL-HH-10K is a large-scale safety dataset with 10,000 high-quality examples, providing near-complete coverage across eight ethical dimensions. It is designed to advance AI safety and ethics research, particularly suitable for safety alignment, interpretable AI, benchmarking, and model training. The dataset is split into training, validation, and test sets, with each example containing detailed annotations and scores, including both rejected and chosen responses. RAIL-HH-10K demonstrates significant improvements in safety and user impact, and offers a multi-dimensional safety scoring tool to assist researchers in evaluating their data.

搜集汇总

背景与挑战

背景概述

RAIL-HH-10K是一个由Responsible AI Labs发布的大规模安全对齐数据集，包含10,000个高质量示例，专注于多维度伦理评估。其核心特点是涵盖公平性、安全性等8个伦理维度，标注覆盖率高达99.5%，每个示例包含有害查询及对应的拒绝与安全响应，并附有详细分数和解释。该数据集适用于AI安全研究、模型训练和基准测试，旨在提升语言模型的多维度安全性能。