Korean Human Preference Dataset

**Overview**<br/>The Human Preference Dataset is a curated Korean-language dataset designed for preference modeling, reward model training, and alignment research. It contains human-annotated preference pairs and evaluations across major knowledge domains. <p><br/></p> **Dataset Composition**<br/>A total of **20,000 Korean-language preference samples**, distributed across three major domain categories: - **Law (Civil, Criminal, Public)** - *6,000 samples*<br/>Includes human preference judgments on legal reasoning, statutory interpretation, case analysis, and domain-specific scenarios. - **STEM (Science, Mathematics, Engineering, Technology)** - *8,000 samples*<br/>Covers problem-solving steps, explanations, reasoning quality, and correctness assessments. - **Economics, Politics, Social Sciences** - *6,000 samples*<br/>Provides preference evaluations on analytical responses, argument quality, policy reasoning, and social context understanding. <p><br/></p> **Key Features** - Fully human-annotated Korean preference pairs - Designed for reward modeling and alignment tasks - Suitable for use in supervised fine-tuning, RM training, or offline RLHF pipelines - Cleaned, deduplicated, and formatted for direct model training <p><br/></p> **Format**<br/>Provided in standard JSON structure typically used for preference datasets.

搜集汇总

背景与挑战

背景概述

这是一个专为偏好建模、奖励模型训练和对齐研究设计的韩语数据集，包含2万个完全由人工标注的偏好样本，覆盖法律、STEM以及经济政治社科三大知识领域。该数据集经过清洗和格式化，适用于监督微调、奖励模型训练等直接模型训练任务。