Erling09/Nemotron-Personas-Korea

Nemotron-Personas-Korea是一个基于韩国真实人口统计、地理和人格特征分布合成的开源人物角色数据集（CC BY 4.0），旨在广泛反映韩国人口的多样性和特征。作为第一个大规模韩语人物角色数据集，它包含姓名、性别、年龄、婚姻状况、教育水平、职业、居住地区等属性，这些属性基于韩国统计厅（KOSIS）、大法院、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。该数据集支持韩国模型开发者构建融入重要区域特定人口统计和文化背景的Sovereign AI系统，可用于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。与现有人物角色数据集相比，它更忠实于真实人口分布，涵盖年龄（如老年人口）、地区（如农村）、教育水平、职业等多个维度。数据集使用NeMo Data Designer（企业级合成数据生成复合AI系统）创建，利用专有概率图模型、Apache-2.0许可的google/gemma-4-31B-it模型以及Data Designer中包含的验证和评估方法。数据集包含100万条记录，涵盖700万个人物角色，26个字段，包括7个人物角色字段、6个人物角色属性字段、12个人口统计和地理上下文字段以及1个唯一标识符，覆盖17个省份和252个地区，包含209,000个唯一姓名（118个姓氏，21,400个名字）和7种人物角色类型（职业、体育、艺术、旅行、烹饪、家庭、简洁）。数据集遵循CC BY 4.0许可，可用于商业和非商业用途。Nemotron-Personas-Korea is an open-source persona dataset (CC BY 4.0) synthesized based on real-world demographic, geographic, and personality trait distributions of South Korea. It is designed to broadly reflect the diversity and characteristics of the South Korean population. As the first large-scale Korean-language persona dataset, it includes attributes such as name, sex, age, marital status, education level, occupation, and region of residence, all synthesized using official statistics from the Korean Statistical Information Service (KOSIS), the Supreme Court of Korea, the National Health Insurance Service, the Korea Rural Economic Institute, and NAVER Cloud. The dataset supports South Korean model builders in developing Sovereign AI systems that incorporate important region-specific demographics and cultural context. It can be used to expand the diversity of synthetic data for sovereign AI model development, mitigate data and model bias, and improve the diversity of model responses. Compared to existing persona datasets, it is designed to more faithfully reflect real population distributions across multiple dimensions, including age (e.g., elderly populations), region (e.g., rural areas), education level, and occupation. The dataset was created using NeMo Data Designer, an enterprise-grade compound AI system for synthetic data generation, leveraging a proprietary probabilistic graphical model (PGM), the Apache-2.0 licensed google/gemma-4-31B-it model, and validation and evaluation methods included in Data Designer. It contains 1M records with 7M personas, 26 fields including 7 persona fields, 6 persona attribute fields, 12 demographic & geographic contextual fields, and 1 unique identifier, covering 17 provinces and 252 districts, with 209K unique names (118 surnames, 21.4K given names) and 7 persona types (professional, sports, arts, travel, culinary, family, concise). The dataset is freely available for both commercial and non-commercial use under CC BY 4.0.