ship-tracking-data
收藏资源简介:
CValues-Comparison是一个专门用于评估和比较大型语言模型在中文场景下安全性与价值观对齐能力的数据集。该数据集旨在为中文大模型的安全和价值对齐研究提供标准化的评估基准,包含两个核心组成部分:1) 安全性及价值观对齐测试数据,约11,000个测试样本,覆盖中文语境下多个维度的安全和价值观问题;2) 人类标注的偏好对,约26,000对标注数据,反映人类对不同模型回复的偏好判断。数据以JSON格式组织,主要字段包括prompt(用户查询)、response(模型回复)、category(问题类别)、safety_label(安全性标签)和value_label(价值观标签),来源于真实的中文用户查询,涵盖安全性和价值观两个关键评估维度。该数据集适用于多种任务场景,如模型安全性评估、价值观对齐度测量、基于人类反馈的偏好学习、红队测试和中文大模型的综合能力评测,为研究人员和开发者提供了一个全面评估中文大模型安全与价值观对齐能力的工具。
CValues-Comparison is a dedicated dataset for evaluating and comparing the safety and value alignment capabilities of large language models (LLMs) in Chinese contexts. It serves as a standardized evaluation benchmark for research on safety and value alignment of Chinese LLMs. The dataset comprises two core components: 1) Safety and value alignment test data, which contains approximately 11,000 test samples covering multiple dimensions of safety and value alignment issues in the Chinese linguistic context; 2) Human-annotated preference pairs, with around 26,000 annotated pairs that reflect human preference judgments regarding responses from different models. The data is structured in JSON format, with core fields including "prompt" (user query), "response" (model response), "category" (question category), "safety_label" (safety label), "value_label" (value alignment label), and others. The data is sourced from real Chinese user queries and encompasses two critical evaluation dimensions: safety and value alignment. This dataset supports a variety of task scenarios, including model safety assessment, value alignment measurement, human feedback-based preference learning, red team testing, and comprehensive capability evaluation of Chinese LLMs. By providing standardized test data and human preference annotations, CValues-Comparison offers researchers and developers a comprehensive tool to evaluate the safety and value alignment capabilities of Chinese LLMs.
数据集概述
- 数据集名称:ship-tracking-data
- 许可证:MIT 协议
此数据集页面仅提供了许可证信息(MIT),未包含数据集的具体描述、用途、样本数量、特征字段等详细内容。




