Nemotron-Personas-USA-Extended

A compound AI approach to personas grounded in real-world distributions

Dataset Overview

Nemotron-Personas-USA-Extended is a commercially-permissive dataset of synthetically-generated personas grounded in real-world demographic, geographic, and personality trait distributions to capture the diversity and richness of the population. It is the first dataset of its kind aligned with statistics for names, sex, age, background, marital status, education, occupation and location, among other attributes. With an initial release focused on the United States, this dataset provides high-quality personas for a variety of modeling use-cases.

The dataset can be used to improve diversity of synthetically generated data, mitigate data and model biases, and prevent model collapse. In particular, the dataset is designed to be more representative of underlying demographic distributions along multiple axes, including age (e.g., older personas), geography (e.g., rural personas), education, occupation and ethnicity, as compared to past Persona datasets.

Produced using NeMo Data Designer, an enterprise-grade compound AI system for synthetic data generation, the dataset leverages a proprietary Probabilistic Graphical Model (PGM) along with an Apache-2.0-licensed GPT-OSS-120B model and an ever-expanding set of validators and evaluators built into Data Designer. It is available for use in NeMo Data Designer itself.

This dataset is ready for commercial/non-commercial use.

What is NOT in the dataset
All data, while mirroring real-world distributions, is completely artificially generated. Any similarity in names or persona descriptions to actual persons, living or dead, is purely coincidental.

Data Developer

NVIDIA Corporation

Release Date

Hugging Face 06/09/2025 via https://huggingface.co/datasets/nvidia/Nemotron-Personas

Dataset Creation Date

06/09/2025

License/Terms of Use

This dataset is licensed under the NVIDIA Dataset License Agreement

Use Case
Developers working on Sovereign AI, training LLMs, and/or looking to improve diversity of synthetically generated data, mitigate data/model biases, and prevent model collapse.

Data Version
1.0 (06/09/2025)
1.1 (10/2/2025)

Intended use

The Nemotron-Personas-USA-Extended dataset is intended to be used by the community to continue to improve open models and push the state of the art. We welcome feedback from the open-source community and invite developers, researchers, and data enthusiasts to explore the dataset and build upon it.

The Nemotron-Personas-USA-Extended dataset is grounded in distributions of self-reported demographic data in the US Census. As such, its primary goal is to combat missing data and/or potential biases present in model training data today, especially when it comes to existing persona datasets used in synthetic data generation. Despite the improved data diversity and fidelity to the US population, we are still limited by data availability and reasonable model complexity. This results in some necessary independence assumptions; for instance, that occupations are independent of location (zip code) given education, age and sex. Similarly, comprehensive statistics on gender, independent of sex, are not available from the Census Bureau. We leave further efforts to improve fidelity to future work.

Note that the dataset is focused on adults only.

Dataset Details

The dataset contains:

• 9M personas across 1M records (9 persona fields & 22 contextual fields)
• ~ 1.6B tokens, including ~816M persona tokens
• 29k geographic areas (ZCTAs) and 15.2k cities across all 50 states + Puerto Rico and Virgin Islands
• 970k unique full names
• 560+ professional occupations, all grounded in real-world distributions
• Comprehensive coverage across demographic and personality trait distributions

Seed Data

In order to capture the socio-demographic and geographic diversity and complexity of the US population, Nemotron-Personas-USA-Extended leveraged open-source (CC0-licensed) aggregated statistical data from:

• The US Census Bureau, specifically the American Community Survey.
• The study “Race and ethnicity data for first, middle, and surnames,” Rosenman et al. (2023); resulting in the dataset located here.

Schema

The dataset includes 22 fields: 6 persona fields and 16 contextual fields shown below. Researchers will find many contextual fields useful in zoning in on specific personas, which is challenging to do with existing datasets.

Nemotron-Personas-USA-Extended
|-- uuid: string Globally unique identifier
|-- professional_persona: string Professional persona capturing primary field of work, key professional skills, traits and behavior
|-- finance_persona: string Financial persona describing spending, saving and investment habits, approach to financial decision-making
|-- healthcare_persona: string Healthcare persona capturing health conditions and approach to medical care
|-- sports_persona: string Sports persona describing athletic interests, sport team affiliations, and approach to fitness and exercise
|-- arts_persona: string Arts persona characterizing engagement with creative expression and how the arts shape their identity
|-- travel_persona: string Travel persona capturing capturing travel interests and style
|-- culinary_persona: string Culinary persona describing food/cuisine preferences, cooking skill level, and approach to dining experiences
|-- persona: string A concise general-purpose persona capturing the essence of a person's perspective and approach to life
|-- detailed_persona: string A detailed persona description encompassing multiple facets of a synthetic individual
|-- openness: string Score, label and description of the openness component in the OCEAN framework
|-- conscientiousness: string Score, label and description of the conscientiousness component in the OCEAN framework
|-- extraversion: string Score, label and description of the extraversion component in the OCEAN framework
|-- agreeableness: string Score, label and description of the agreeableness component in the OCEAN framework
|-- neuroticism: string Score, label and description of the neuroticism component in the OCEAN framework
|-- cultural_background Description of the person's cultural background |-- skills_and_expertise: string Professional and personal skills in narrative format
|-- career_goals_and_ambitions: string Professional aspirations and long-term career objectives
|-- first_name First name of the synthetic individual
|-- last_name Last name of the synthetic individual
|-- hobbies_and_interests: string Personal interests and recreational activities in narrative format
|-- skills_and_expertise_list: string List of skills and areas of expertise
|-- hobbies_and_interests_list: string List of hobbies and personal interests
|-- sex: string Biological sex (e.g., Male, Female)
|-- age: integer Age in years
|-- marital_status: string Relationship status (e.g., currently married, never married, divorced, widowed, separated)
|-- education_level: string Highest level of education completed
|-- bachelors_field: string Field of study for bachelor's degree, if applicable
|-- occupation: string Comprehensive professional occupation
|-- city: string City of residence
|-- state: string State of residence
|-- zipcode: string Postal code of residence
|-- country: string Country of residence

Field & Token Counts

1.6B tokens across 1M records and 22 columns, excluding the globally unique identifier. Note that data covers 50 states as well as Puerto Rico and Virgin Islands.

Dataset Description & Quality Assessment

The analysis below provides a breakdown across various axes of the dataset to emphasize the built-in diversity and pattern complexity of data.

Names

Since the focus of this dataset is on personas, names aren’t provided as dedicated fields. However, infused into persona prompts are 136,000 unique first_names, 126,000 unique middle names, and 338,000 unique surnames sourced from Rosenman et al. (2023).

Age distribution

The plots below provide an analysis of the demographic sampling procedure used to seed personas (using 1M samples).

The distribution of our persona ages takes the form of a bulging population pyramid that reflects historical birth rates, mortality trends, and migration patterns. This is in stark contrast to a bell curve distribution typically produced by an LLM alone. Overall the distribution is right-skewed and distinctly non-Gaussian. Note that minors are excluded from this dataset (see the Ethics section below).

Marital Status by Age Group

The heatmap below displays the fraction of people for each age cohort who are (1) never married, (2) currently married, (3) separated, (4) divorced, or (5) widowed. It highlights how marital status shifts over the life course in the US with “never married” dominating late teens and early twenties, “married” climbing rapidly in twenties and peaking in mid-fourties, divorced and widowed being much more pronounced in later stages of life. All of these considerations are of relevance to informing life experiences and personas.

Education Level by Age Group

The heatmap below captures intricate patterns of educational attainment across age cohorts. For example, note how the share of high-school-only and no-diploma individuals ebbs then resurges among the oldest age groups, reflecting historical shifts in access and in social norms.

Geographic Intricacies of Education Attainment

This slice of our dataset demonstrates how geography informs education and therefore persona descriptions. The choropleth map shows, for each U.S. state, the share of residents age 25 and older who hold at least a bachelor’s degree. No LLM in our testing was able to generate data of this fidelity.

| ----- | ----- | | Nemotron-Personas-USA-Extended dataset | Source: Wikipedia, Educational attainment in the United States, Wikipedia |

Occupational Categories

The treemap below reflects the richness of our dataset with respect to professional occupations of personas. Represented in our dataset are over 560 occupation categories that are further informed by demographic and geographic distributions.

Persona diversity

The attributes above (and many more) ultimately affect the diversity of the synthetic personas being generated. As an example, the analysis below highlights a multitude of clusters within professional persona descriptions. These clusters are identified by clustering embeddings and reducing dimensionality to 2D.

How to use it

You can download the dataset from NGC or use it directly in Data Designer as follows:

config_builder = DataDesignerConfigBuilder(model_configs=model_configs)
# USA personas in English (change the en_US locale for other languages) personas
config_builder.add_column(
   SamplerColumnConfig(
       name="person",
       sampler_type=SamplerType.PERSON,
       params=PersonSamplerParams(
           locale="en_US",
           age_range=[18, 114],
       ),
   )
)

Dataset Characterization

Data Collection Method

• Hybrid: Human, Synthetic, Automated

Labeling Method

• Not Applicable

Dataset Format

• Text

Dataset Quantification

• Record counts: 1M records (6M persona descriptions)
• Total data storage: 2.6GB

Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report security vulnerabilities or NVIDIA AI concerns here.

Citation

If you find the data useful, please cite:

@software{nvidia/Nemotron-Personas-USA,
  author = {Meyer, Yev and Corneil, Dane},
  title = {{Nemotron-Personas-USA}: First Synthetic Personas Dataset Aligned to Real-World Distributions
},
  month = {June},
  year = {2025},
  url = {https://huggingface.co/datasets/nvidia/Nemotron-Personas-USA}
}