NVIDIA Releases Nine Million Synthetic Japanese Financial Personas

Elith and NVIDIA released Nemotron-Personas-Finance-Japan on November 9, 2026, opening nine million synthetic Japanese personas — one million Japanese-language records with nine personas each — for commercial and non-commercial AI development under a CC BY 4.0 license. The dataset adds household finances, money behavior and financial literacy to demographic profiles, giving banks, insurers and developers a way to test customer-facing systems without touching real customer records.

The release matters because Japanese financial institutions face a market where people aged 65 or older make up 29.3% of the population as of October 1, 2024, according to the Statistics Bureau's 2024 population estimates. Retirement planning, asset succession and inheritance workflows now define a large share of customer interactions, and no open dataset at this scale has combined Japan's official statistics with household-finance and financial-literacy data — a claim NVIDIA itself qualifies as true "to its knowledge."

What does the new dataset contain?

The dataset extends Nemotron-Personas-Japan, whose September 23, 2025 version held six million personas across 22 fields and excluded finance-specific personas. The new edition contains 50 fields per record, covering all 47 prefectures and more than 1,600 occupation categories. It covers only adults aged 18 and older.

Three of the 20 persona-related fields separately describe financial habits, decision-making behavior and financial concerns. The 30 contextual fields add:

• personal income and household assets
• housing and mortgage status
• consumption propensity
• financial-literacy scores
• internet and smartphone use
• household structure
• self-rated health

Those profiles can seed multi-turn Japanese dialogue, from a regional pensioner asking about asset succession to an urban dual-income household reviewing education savings and a mortgage.

Why do the household-asset numbers matter?

The household-asset distributions draw on the Statistics Bureau's 2019 National Survey of Family Income, Consumption and Wealth, which investigates family consumption, income, assets and liabilities. The dataset card's analysis reports that the share of households with at least 20 million yen in financial assets rises from about 10% at ages 25–29 to about 30% at ages 60–64.

Assets remain relatively high through people's 60s and 70s even as personal income falls around retirement. That pattern supports distinct customer profiles for wealth building, retirement planning and inheritance — the exact segments where Japanese banks and insurers are concentrating AI-assisted service development.

The demographic context sharpens the point. The Statistics Bureau's 2024 population estimates put people aged 65 or older at 29.3% of Japan's population as of October 1, 2024, making retirement and inheritance simulations a primary intended use case.

How were the personas built?

The generation stack combines three components:

• NeMo Data Designer, which supplies structured outputs, validation and automated retries
• the NeMo SDG-PGMs probabilistic-graphical-model framework (a method that encodes statistical relationships between variables to generate consistent synthetic data)
• the Apache-2.0-licensed gemma-4-31B-it model for Japanese text

Sompo Japan, Tokio Marine Holdings, Bank of Fukuoka, Mizuho Financial Group, Meiji Yasuda Life Insurance and the Financial Data Utilization Association provided feedback during the design process. That roster of six major Japanese financial institutions signals direct industry demand for testable synthetic customers rather than real customer records.

The dataset card names J-FLEC's survey and statistical data as another seed input. J-FLEC's official index says its financial-literacy survey covers people aged 18–79, while its household financial-behavior surveys examine assets, liabilities and life planning for two-person and single-person households.

What are the privacy and disclosure caveats?

Every persona is artificial rather than a record of a real customer or individual, and any resemblance to an actual person is coincidental, according to the dataset documentation. The public Hugging Face version omits standalone first- and last-name fields and synthetic addresses, and excludes healthcare personas.

The card also discloses a statistical assumption: occupation is independent of education given prefecture and sex. It notes that comprehensive census statistics on gender independent of sex are unavailable. Both caveats matter for developers who need to understand where the synthetic distributions may diverge from real population data.

Intended uses listed for the personas include customer-service agents, sales and marketing conversation tests, NeMo User Sim simulations, red-teaming (adversarial testing of model safety) and checks for consistent behavior across rural and urban locations, ages, education levels and economic circumstances.

What is the companion PII dataset?

NVIDIA paired the release with Nemotron-PII-Japan, a companion dataset for training and testing systems that detect and redact personally identifiable information. It contains 100,000 synthetic Japanese business documents spanning 43 document types in eight domains, with 56 PII labels and 810,703 annotated spans.

The two datasets are designed to work together. The finance personas simulate customers while the PII documents test whether a pipeline identifies sensitive data — the two failure modes, poor interaction quality and privacy leakage, that financial AI teams must control simultaneously.

The PII dataset card carries an explicit caution: good results on synthetic documents do not guarantee complete detection, fairness, privacy protection, regulatory compliance or prevention of leakage on real data.

What comes next?

The releases give financial AI teams an open way to test both customer interactions and privacy controls in Japanese, at a scale — nine million personas and 100,000 documents — that was previously unavailable under an open license. Real-world performance against institutions' own Japanese-language workflows, not benchmark results on the synthetic data itself, is therefore the decisive measure of the releases' practical value.