Nvidia and Elith release a 9 million persona Japanese finance dataset under an open license

NVIDIA and Elith have released Nemotron-Personas-Finance-Japan, a 9-million-persona synthetic dataset built on Japan's official demographic and labor statistics. The dataset, available under a CC BY 4.0 license for commercial use, gives financial institutions and developers a privacy-safe foundation for building and testing AI systems without touching real customer data.

What's inside the dataset

Each of the 1 million base records expands into nine distinct personas, totaling 9 million individual profiles written in natural Japanese. The records span 50 fields - including nine persona fields, 11 persona attributes, and 30 contextual fields tied to government statistics. The dataset contains roughly 1.5 billion tokens, with 750 million of those dedicated to persona descriptions.

Coverage runs across demographic, regional, asset distribution, and personality trait axes. The 950,000 unique names and 1,600-plus occupation categories reflect Japan's actual workforce distribution. Persona types include occupations, sports, arts, travel, cooking, and finance, with natural-language attributes describing cultural background, skills, career goals, household finances, attitudes toward money, and financial planning.

How the data was generated

The team used NVIDIA's NeMo Data Designer, a compound AI system that handles complex Jinja templating, Pydantic validation, structured outputs, and automated retries across multiple generation backends. Statistical grounding came from NeMo SDG-PGMs, a probabilistic graphical model framework released under Apache 2.0. Japanese text generation ran on gemma-4-31B-it, also Apache 2.0 licensed.

Several design choices push the dataset beyond raw statistical distributions. Education fields carry finer granularity than national aggregates. Occupation categories include business owners and specialized professions beyond standard classifications. Life-stage modeling covers students, retirees, and periods of unemployment - scenarios that national statistics often underrepresent. Household-level financial and real assets sit alongside individual income and spending data. Internet and smartphone usage metrics pair with financial literacy indicators to capture consumer profiles relevant to financial use cases.

Built with industry input

The dataset was designed with feedback from major Japanese financial institutions and industry groups, including Sompo Japan, Tokio Marine Holdings, Bank of Fukuoka, Mizuho Financial Group, Meiji Yasuda Life Insurance, and the Financial Data Utilization Association. Every persona is fully synthetic - demographics and household finances follow official statistics and public survey data, and names follow real Japanese naming patterns, but no records correspond to actual individuals.

Practical applications

Teams can use the personas as seeds for multi-turn conversation generation, producing human-like dialogue datasets for customer-service agents. The dataset supports domain-specific AI development for financial assistants, bias testing across rural-urban divides and age groups, and fairness evaluations that check whether models perform consistently across Japanese society. The personas integrate with NeMo User Sim, enabling simulations drawn from distributions that mirror real-world populations.

The release pairs with Nemotron-PII-Japan, a same-day launch containing 100,000 synthetic Japanese business documents with labeled personal information. Use the finance personas to simulate customers, and the PII dataset to test how pipelines protect their data.

Why this matters for finance, insurance, and product teams

For teams building customer-facing AI in regulated markets, synthetic data eliminates the compliance bottleneck of using real customer records during development. The 50-field structure - covering household assets, financial literacy, spending propensity, and digital behavior - lets product teams segment populations and test model behavior across economic strata before deployment. Insurance underwriters and marketing teams can stress-test personalization engines against demographic slices that reflect Japan's actual population distribution, not an approximated Western dataset. The CC BY 4.0 license means commercial teams can integrate the data directly into production pipelines without licensing friction.