fusion-dw
收藏资源简介:
CValues-Comparison是一个专门设计用于评估大型语言模型在中文语境下价值观对齐能力的数据集。该数据集围绕中文价值观的核心维度构建,包含安全、责任和道德三个主要方面。数据集包含约1000个精心设计的中文问题,每个问题都配有4-5个不同的回答选项,这些选项反映了在特定情境下不同的价值观取向和行为选择。数据来源于基于人类价值观的标注工作,旨在捕捉中文文化和社会背景下的价值观多样性。该数据集主要用于价值观对齐评估、模型安全评估和不同模型之间的比较分析。通过分析模型对不同回答选项的偏好和选择模式,研究人员可以量化评估模型与人类价值观的一致性程度,为模型的安全部署和价值观对齐优化提供重要参考。使用该数据集时,建议采用标准化的评估流程以确保评估结果的公平性和可比性。
CValues-Comparison is a dataset specifically designed to evaluate the value alignment capabilities of large language models in Chinese contexts. The dataset is constructed around core dimensions of Chinese values, including safety, responsibility, and morality. It contains approximately 1,000 carefully designed Chinese questions, each accompanied by 4-5 different answer options that reflect different value orientations and behavioral choices in specific scenarios. The data is sourced from human value-based annotation efforts, aiming to capture the diversity of values within Chinese cultural and social contexts. This dataset is primarily used for value alignment assessment, model safety evaluation, and comparative analysis between different models. By analyzing the preferences and selection patterns of models across different answer options, researchers can quantitatively assess the consistency of models with human values, providing important references for safe model deployment and value alignment optimization. When using this dataset, it is recommended to follow standardized evaluation procedures to ensure the fairness and comparability of assessment results.
数据集概述
- 数据集名称:fusion-dw
- 数据集来源:Hugging Face Datasets,链接为:https://huggingface.co/datasets/StringFellow/fusion-dw
- 许可证:Apache-2.0
该数据集使用Apache 2.0开源许可证,除此之外未提供其他描述信息或数据内容说明。




