Nemotron-Personas-Germany

Nemotron-Personas-Germany 是由 Soofi 与 NVIDIA 发布的德语开源合成人物（persona）数据集，采用 CC BY 4.0 许可，可用于商业与非商业用途。其目标是基于德国真实人口与地理特征分布合成具有多样性和代表性的人物画像，支持主权 AI 与德语 LLM 开发，提升合成数据多样性、减轻数据与模型偏差，并帮助防止模型崩溃。数据依据德国 2022 年人口普查（Zensus 2022）及德国联邦统计局（Destatis）和各州统计局发布的官方统计合成，并与德国 AI 研究机构和初创企业联盟 Soofi 合作完成。它是首个此类德语数据集，对齐姓名、性别、年龄、移民背景、婚姻状况、家庭类型、教育、就业、职业和居住地等统计维度。数据集使用 NVIDIA NeMo Data Designer 生成，结合专有概率图模型（PGM）、Apache-2.0 许可的 google/gemma-4-31B-it 模型以及内置验证器和评估器。发布版本为 1.0，发布于 2026 年 10 月 20 日，共 1,000,000 条训练记录，含 600 万个德语 persona；约 12.4 亿 tokens，其中 persona tokens 约 4.33 亿。数据覆盖德国 16 个联邦州和 400 个县及县级市，包含 775,000 个唯一完整姓名（15,835 个唯一名、80,853 个唯一姓）、约 12,000 个职业（20,000 个性别特定职称）和 1,243 个 KldB 2010 职业类型。YAML 元数据将任务类别标为 text-generation，语言为 de，标签包括 synthetic、personas、NVIDIA、German、datadesigner，规模类别为 1M<n<10M；train split 为 1,000,000 条，dataset_size 为 4,622,057,563 bytes，download_size 为 1,931,953,520 bytes。数据集包含 27 个字段：唯一标识 uuid；6 个 persona 字段（professional_persona、sports_persona、arts_persona、travel_persona、culinary_persona、persona）；6 个自然语言 persona 属性字段（cultural_background、skills_and_expertise、skills_and_expertise_list、hobbies_and_interests、hobbies_and_interests_list、career_goals_and_ambitions）；14 个基于官方人口与劳动统计的上下文与人口字段（sex、age、marital_status、migration_background、household_type、housing_tenure、education_level、employment_status、job_position、industry、occupation、kreis、bundesland、country）。其中 age 为 int64，其余字段通常为字符串。数据在区域、移民背景、家庭构成、教育水平和职业等方面力求比以往 persona 数据集更具代表性，可用于生成包含真实姓名、年龄和背景的高质量多轮聊天对话等合成数据。README 明确指出数据集仅面向成人（18 岁及以上），所有数据虽反映真实分布但完全人工生成，与真实人物相似纯属巧合。数据集不包含 NeMo Data Designer 中其他可用字段，如独立的名或姓、合成地址、健康与金融 persona、Big Five 人格特质等。KldB 职业代码本身不在发布字段中，发布的是职业名称（occupation）和行业（industry）。数据基于 Zensus 2022、Destatis、各州统计局以及 KldB 2010 等种子数据；统计性别身份（独立于生理性别）的细粒度官方数据不足，名称是统计近似而非普查。目标用户包括从事主权 AI、LLM 训练、合成数据生成、偏差缓解和模型崩溃预防的开发者与研究者。Nemotron-Personas-Germany is an open-source German synthetic persona dataset released by Soofi and NVIDIA under the CC BY 4.0 license, available for commercial and non-commercial use. Its goal is to synthesize diverse and representative persona profiles based on real German population and geographic distributions, supporting sovereign AI and German LLM development, improving synthetic data diversity, reducing data and model bias, and helping prevent model collapse. The data is synthesized based on Germanys 2022 Census (Zensus 2022) and official statistics published by the Federal Statistical Office of Germany (Destatis) and state statistical offices, and was completed in collaboration with Soofi, an alliance of German AI research institutions and startups. It is the first such German-language dataset, aligning statistical dimensions such as name, sex, age, migration background, marital status, household type, education, employment, occupation, and place of residence. The dataset was generated using NVIDIA NeMo Data Designer, combining a proprietary probabilistic graphical model (PGM), the Apache-2.0 licensed google/gemma-4-31B-it model, and built-in validators and evaluators. Release 1.0, published on October 20, 2026, contains 1,000,000 training records and 6 million German personas; it has about 1.24 billion tokens, of which approximately 433 million are persona tokens. The data covers 16 German federal states and 400 counties and independent cities, and includes 775,000 unique full names (15,835 unique first names and 80,853 unique last names), about 12,000 occupations (20,000 gender-specific job titles), and 1,243 KldB 2010 occupation types. YAML metadata marks the task category as text-generation, the language as de, and tags including synthetic, personas, NVIDIA, German, and datadesigner; the size category is 1M<n<10M. The train split contains 1,000,000 records, dataset_size is 4,622,057,563 bytes, and download_size is 1,931,953,520 bytes. The dataset contains 27 fields: the unique identifier uuid; 6 persona fields (professional_persona, sports_persona, arts_persona, travel_persona, culinary_persona, persona); 6 natural-language persona attribute fields (cultural_background, skills_and_expertise, skills_and_expertise_list, hobbies_and_interests, hobbies_and_interests_list, career_goals_and_ambitions); and 14 context and demographic fields based on official population and labor statistics (sex, age, marital_status, migration_background, household_type, housing_tenure, education_level, employment_status, job_position, industry, occupation, kreis, bundesland, country). Among these, age is int64, and the other fields are usually strings. The data strives to be more representative than previous persona datasets in terms of region, migration background, household composition, education level, and occupation, and can be used to generate synthetic data such as high-quality multi-turn chat conversations containing realistic names, ages, and backgrounds. The README explicitly states that the dataset is intended only for adults aged 18 and above; although all data reflects real distributions, it is entirely artificially generated, and any resemblance to real persons is purely coincidental. The dataset does not include other fields available in NeMo Data Designer, such as separate first or last names, synthetic addresses, health and financial personas, or Big Five personality traits. The KldB occupation codes themselves are not among the released fields; what is released is the occupation name (occupation) and industry (industry). The data is based on seed data such as Zensus 2022, Destatis, state statistical offices, and KldB 2010; there is insufficient fine-grained official data on statistical gender identity (independent of biological sex), and names are statistical approximations rather than census-derived. Target users include developers and researchers working on sovereign AI, LLM training, synthetic data generation, bias mitigation, and model collapse prevention.