Nvidia’s 9M synthetic Japanese personas add household finance data

Nvidia, working with Elith, has published Nemotron-Personas-Finance-Japan on Hugging Face: 9 million fully synthetic Japanese personas that add household assets, financial literacy and attitudes toward money to official demographic statistics, under a CC BY 4.0 license.

Nine personas in each of a million records

The dataset holds 1 million records with nine personas each, written in Japanese. Nvidia puts the total at about 1.5 billion tokens, roughly 750 million of them persona text. It lists about 950,000 unique names and more than 1,600 occupation categories.

Each record carries 50 fields: 9 persona fields, 11 persona attributes and 30 contextual fields based on official demographic and labor statistics. The finance layer covers household-level financial and real assets, individual income and spending propensity, financial literacy, internet and smartphone use, and attitudes toward financial planning. There are also persona types for finance, financial behavior and financial concerns.

Nvidia says the data was built from Japanese government statistics with NeMo SDG-PGMs, a probabilistic graphical model framework, while NeMo Data Designer produced the text. The Japanese text came from gemma-4-31B-it, which is licensed under Apache-2.0. According to the post, the design drew on feedback from 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. Nvidia says it refined education, occupation and life-stage categories beyond national statistics, for example by adding finer education levels and business owners.

On privacy, Nvidia states: “Every persona is fully synthetic. No real individuals are represented, and the dataset is designed to reduce privacy risk.” Names follow real Japanese patterns, but the personas are drawn from statistical distributions rather than real records.

Who can use it today, and for what

The dataset is available now at nvidia/Nemotron-Personas-Finance-Japan and loads through the Hugging Face datasets library. Nvidia says the CC BY 4.0 license permits commercial use. Nvidia released a companion set the same day, Nemotron-PII-Japan, with 100,000 synthetic Japanese business documents containing labeled personal information. It also points to NeMo User Sim as a simulation framework for using such personas.

In practice, the data is aimed at teams that cannot use real customer records. Banks, insurers and fintech developers in Japan can test conversational assistants, run simulated customer interviews or build evaluation sets without touching personal data. Model builders can use the Japanese text to tune assistants for financial questions in the local language. Researchers get a distribution-matched population they can share freely, which real customer data rarely allows.

Compared with what existed before, the change is mostly one of depth. The earlier Nemotron-Personas-Japan, released in September 2025, had 1 million records with 6 personas each, or 6 million personas, and 22 fields per record. The new set raises that to 9 million personas and 50 fields, which matches the 50% increase in personas that Nvidia cites. The earlier release was aimed at developers of Japanese sovereign AI systems; this one narrows the focus to finance.

What the announcement leaves unproven

Nvidia calls the dataset the first open persona set at million-record scale to pair official Japanese statistics with household finance and financial literacy data. That is the company’s own claim, and the post as read does not include an independent comparison with other datasets.

More important for users, the post describes how the data was grounded but, in the version we read, gives no measurements of how closely the synthetic financial fields match real Japanese households. Statistical grounding in national surveys does not guarantee that the combined fields, such as asset levels tied to a particular attitude toward risk, behave like real customers. Persona text generated by a language model can also carry that model’s stereotypes about age, region or income. Anyone using the data to evaluate a financial product should check those distributions against their own customer data first.

The post also does not report how much of the refinement beyond official statistics, such as the added education levels and business-owner categories, rests on assumptions rather than published data. The financial institutions named gave feedback on the design, but the post does not say that any of them has adopted the dataset or validated its fields. Finally, a mirror of the earlier dataset on Nvidia’s NGC catalog refers to the NVIDIA Dataset License Agreement alongside CC BY 4.0, so users should confirm the license file on the Hugging Face dataset page before commercial use.