A privacy-preserving, open dataset developed with U.S. Census data
Open Data for American AI Innovation
As the inspiration for AI systems shifts from scraped web text to verifiable, high-quality data, NVIDIA’s Nemotron-Personas-USA dataset provides a transparent, privacy-safe alternative — built entirely from synthetic data validated against U.S. Census distributions.
Created with NVIDIA NeMo Data Designer, the dataset comprises 6 million fully synthetic American personas spanning all 50 states and territories. Each profile reflects realistic demographic, occupational, and behavioral traits designed to mirror the range of the U.S. population — without exposing any personally identifiable information (PII).
What’s within the Dataset
- 6 million personas, aligned with U.S. Census Bureau and BLS statistics
- ~936 million tokens (~371 million persona tokens)
- 970k unique full names (53.7k first | 43.2k middle | 118k last)
- 560+ occupations grounded in real-world distributions
- 18.7k ZCTAs and 9.5k cities across 50 states + territories
- Coverage of underrepresented groups by age, geography, education, and ethnicity
- 100% synthetic → zero PII risk
- License: CC BY 4.0 (open for research and industrial use)
How We Built It
A compound AI approach to personas grounded in real-world distributions
Data Generation Pipeline
Built using NeMo Data Designer, NVIDIA’s compound AI microservice for large-scale synthetic data generation. The system supports Jinja templating, Pydantic validation, structured outputs, automated retries, and multiple generation backends — enabling datasets at national scale.
Core techniques:
- Probabilistic Graphical Models (Apache-2.0) for statistical grounding
- Multi-LLM ensemble for semantic and contextual consistency
- Personality modeling (OCEAN) for behavioral diversity
Private by Design
No real names. No re-identification risk.
All personas are fully synthetic. While grounded in aggregate U.S. statistics, no record is linked to any individual. This ensures developers can safely train AI systems without privacy risks or regulatory barriers — a requirement for trustworthy AI in finance, healthcare, and public sector applications.
Who This Is For
Designed for AI developers, researchers, and policy teams constructing Sovereign AI solutions that reflect U.S. culture and context.
Practical AI Applications
Developers can mix Nemotron-Personas data with other NeMo toolkits to construct end-to-end data pipelines:
- Train and evaluate expert AI agents for finance, healthcare, and public services
- Minimize sensitive data risk when developing AI models or APIs
- Create “what-if” simulations for policy or market forecasting
- Prevent data drift and model collapse through continuous synthetic refresh
Why It Matters
Trustworthy AI relies on trustworthy data.
Traditional anonymization often fails to fulfill modern privacy or reproducibility standards. Against this, fully synthetic datasets offer:
- Provable privacy and compliance — no link to real people
- Transparent provenance — reproducible generation pipeline
- Cultural grounding — aligned to U.S. regional and demographic statistics
- High utility — performance parity with real data on downstream tasks
This release expands NVIDIA’s growing Nemotron-Personas collection, now spanning the U.S., Japan, and India — a foundation for Sovereign AI development and localized model fine-tuning worldwide.
Start Constructing with Nemotron-Personas-USA
To start experimenting today:
from datasets import load_dataset
nemotron_personas_us = load_dataset("nvidia/Nemotron-Personas-USA")
Whether you are developing Sovereign AI applications for U.S. institutions or constructing global agents that require accurate cultural context, Nemotron-Personas-USA provides the authentic, privacy-safe foundation your applications need.
Download it. High quality-tune it. Construct AI that understands American culture and values.
Should you’re able to go deeper, an prolonged version of Nemotron-Personas-USA (including synthetic addresses, occupation hierarchies, and income bands) is accessible through NVIDIA NeMo Data Designer.
