value-for-instruction-tuning
收藏资源简介:
该数据集是一个专为指令微调设计的统一道德价值数据集,由亚琛工业大学和波恩大学等机构的研究者构建,旨在解决大语言模型在特定人类价值观对齐方面的挑战。数据集整合了ETHICS、UNIMORAL和SOCIAL-CHEM-101三个现有高质量道德价值基准,通过数据预处理、生成器填补缺失值以及提示工程转化为统一的指令-响应格式,从而形成一个可直接用于模型训练的语料库。其创建过程涉及将异构的道德场景与判断标准化,并利用训练好的生成器在规范伦理和道德基础理论两种价值框架下进行数据扩展,确保了数据的一致性和丰富性。该数据集主要应用于大语言模型的价值观对齐研究,通过指令微调帮助模型学习价值感知行为,同时保持通用任务性能,为探索数据混合比例如何影响价值导向任务表现提供了重要资源。
This dataset is a unified moral value dataset specifically designed for instruction tuning, constructed by researchers from institutions including RWTH Aachen University and the University of Bonn, aiming to address the challenges in aligning large language models (LLMs) with specific human values. It integrates three existing high-quality moral value benchmarks: ETHICS, UNIMORAL, and SOCIAL-CHEM-101, and is converted into a unified instruction-response format through data preprocessing, missing value filling via generators and prompt engineering, thus forming a corpus directly applicable for model training. The dataset creation process involves standardizing heterogeneous moral scenarios and judgments, and leveraging trained generators to conduct data augmentation under two value frameworks: normative ethics and moral foundations theory, ensuring the consistency and richness of the data. This dataset is primarily utilized in value alignment research for LLMs: it helps models learn value-aware behaviors via instruction tuning while retaining general task performance, serving as a critical resource for exploring how data mixing ratios influence the performance of value-oriented tasks.
数据集概述
- 名称: Value-for-instruction-tuning
- 许可证: CC-BY-SA-4.0
- 来源: 基于三个现有数据集构建:
核心特征
- 对每个场景额外标注了规范伦理学标签和道德基础标签
- 数据已包装为指令调优模板,可直接用于指令调优
外部资源
- 更多信息请参阅 GitHub 仓库

- 1A Unified Moral-Value Dataset for Instruction Tuning亚琛工业大学; 波恩大学; 拉马尔机器学习和人工智能研究所 · 2026年



