Synthetic training data can inherit a model’s narrow view of a population. This dataset tackles that problem by anchoring persona generation in Germany’s 2022 census and labor statistics, so demographic and regional variety is part of the seed data rather than an afterthought.
What Sets It Apart
- One million records provide six persona descriptions per record, covering professional, sports, arts, travel, culinary, and concise general-purpose perspectives.
- Demographic context spans adults, all 16 federal states, and 400 districts. This makes it possible to condition downstream data generation on attributes such as age, education, employment, and region.
- A probabilistic graphical model supplies statistically grounded profiles; NeMo Data Designer and Gemma 4 31B generate the natural-language attributes and persona descriptions. The result is reusable seed material for creating more varied training conversations.
- The records are synthetic, not census microdata. Names and descriptions are generated, and the dataset is released under CC BY 4.0 for commercial and non-commercial use.
Who It’s For
A good fit for teams building German-language models or synthetic conversation data that need better demographic and regional coverage. Look elsewhere if you need real individual records, profiles of people under 18, dedicated healthcare or finance personas, or a dataset that directly provides names as separate fields. Its fidelity is also bounded by the available statistics and the assumptions used to model them.
