Nemotron-Personas-Finance-Japan: 9 million synthetic personas for financial AI

The problem the dataset tries to solve is privacy. In the financial sector, real data are hard to use because of regulations, compliance and the need for domain expertise. Without representative data, financial AI agents are trained and evaluated in artificial scenarios.

In this article you will see:

• What is inside the dataset and how it was built
• Why household finances and financial literacy matter
• Practical use cases, from support to red teaming
• How to test today, with code and license

What is inside Nemotron-Personas-Finance-Japan

The dataset has 9 million personas in total: 1 million records with 9 personas each. Each record has 50 fields, with 9 persona fields, 11 persona attributes and 30 contextual fields based on official demographic and labor statistics.

In total there are about 1.5 billion tokens, including 750 million persona tokens. There are about 950 thousand unique names and more than 1,600 occupation categories that reflect the Japanese workforce. Coverage spans demographic axes, regional, asset distribution and personality traits.

Persona types include occupations, sports, arts, travel, cooking, finance, financial behavior and financial concerns. Natural language attributes include cultural background, skills, career goals, hobbies, family and financial context, attitudes toward money and financial planning.

How the dataset was built

Construction uses the NeMo SDG-PGMs structure, a probabilistic graphical model framework for statistically grounded synthetic data generation, and NeMo Data Designer, an enterprise synthetic data generation system. NeMo Data Designer allows complex Jinja templates, Pydantic validation, structured outputs, automatic retries and support for multiple backends.

For Japanese text generation, the pipeline uses the gemma-4-31B-it model, licensed under Apache 2.0. Demographic and financial data follow official Japanese government statistics and public surveys.

An important point is cultural refinement. Where national statistics aggregate education levels broadly, the dataset introduces finer granularity. Occupations include additional categories, such as business owners and specialized professions. Life stages model scenarios rarely highlighted in statistics, such as students, retirees and periods of unemployment. Household finances include financial and real assets at household level, in addition to individual income.

Why household finances and financial literacy matter

The dataset was designed with participation from Japanese financial institutions and industry associations, including Sompo Japan, Tokio Marine Holdings, Bank of Fukuoka, Mizuho Financial Group, Meiji Yasuda Life Insurance and the Financial Data Utilization Association. This collaboration ensures financial scenarios are realistic and not just demographically plausible.

Each persona carries household finances and financial literacy along with demographics. Teams can use the personas to build customer service agents, evaluate sales and marketing conversations, run user simulations in NeMo User Sim and do red teaming of customer facing AI.

The dataset is 50% larger than the original Nemotron-Personas-Japan, with 50 fields instead of 22, offering finer population slices to train and test.

Privacy by design

All personas are fully synthetic. No real individual is represented. Demographics and household finances follow official statistics and public research data. Names follow real Japanese naming patterns, but personas are built from those distributions, not from records about real people.

This allows teams to develop and evaluate financial AI without using real customer data, reducing privacy risk and data acquisition cost.

Practical use cases

The dataset supports applications like multi-turn dialogue generation using personas as seeds, domain-specific AI development for financial assistants, and bias and fairness testing evaluating how models perform across rural-urban divides, age groups, educational backgrounds and economic circumstances.

The release also integrates with the NeMo User Sim user simulation framework, enabling instant execution of simulations from diverse distributions consistent with the real world.