an2zzang/Nemotron-Personas-Korea

Nemotron-Personas-Korea 是一个基于韩国真实人口统计、地理和性格特征分布合成的开源人物角色数据集，采用CC BY 4.0许可证。该数据集设计用于广泛反映韩国人口的多样性和特征，是首个大规模韩语人物角色数据集。它包含100万条记录，涵盖700万个人物角色，具有26个字段，包括7个人物角色字段（如职业、体育、艺术、旅行、烹饪、家庭和简洁类型）、6个人物角色属性字段（如文化背景、技能与专长、职业目标与抱负、爱好与兴趣列表）和12个人口统计与地理上下文字段（如性别、年龄、婚姻状况、教育水平、职业、居住地区等），以及一个唯一标识符。数据集基于韩国统计信息服务、最高法院、国民健康保险公团、农村经济研究院和NAVER Cloud的官方统计数据合成，旨在支持韩国模型开发者构建包含区域特定人口统计和文化背景的主权AI系统，以扩大合成数据的多样性、减轻数据和模型偏见，并提高模型响应的多样性。数据集使用NeMo Data Designer创建，采用概率图模型和gemma-4-31B-it模型生成，并经过验证和评估。数据集仅包含成年人物角色（19岁及以上），可用于商业和非商业用途。Nemotron-Personas-Korea is an open-source persona dataset synthesized based on real-world demographic, geographic, and personality trait distributions of South Korea, licensed under CC BY 4.0. It is designed to broadly reflect the diversity and characteristics of the South Korean population and is the first large-scale Korean-language persona dataset. The dataset contains 1 million records covering 7 million personas, with 26 fields including 7 persona fields (e.g., professional, sports, arts, travel, culinary, family, concise), 6 persona attribute fields (e.g., cultural background, skills and expertise, career goals and ambitions, hobbies and interests lists), and 12 demographic and geographic contextual fields (e.g., sex, age, marital status, education level, occupation, region of residence), along with a unique identifier. It is synthesized using official statistics from the Korean Statistical Information Service, Supreme Court, National Health Insurance Service, Korea Rural Economic Institute, and NAVER Cloud. The dataset supports South Korean model builders in developing Sovereign AI systems that incorporate region-specific demographics and cultural context, aiming to expand the diversity of synthetic data, mitigate data and model biases, and improve model response diversity. Created using NeMo Data Designer with a probabilistic graphical model and the gemma-4-31B-it model, it includes validation and evaluation methods. The dataset only includes adult personas (19 years and older) and is freely available for both commercial and non-commercial use.