NVIDIA and Elith release Nemotron-Personas-Finance-Japan, 9M synthetic personas

Elith and NVIDIA have released Nemotron-Personas-Finance-Japan, a financial-domain extension of the open-source Nemotron-Personas-Japan dataset. The pitch is simple: banks, insurers and their suppliers in Japan face strict privacy rules and compliance demands when they try to use real customer data, so AI work for finance needs synthetic data that still reflects real demographics, household structures and asset positions. The authors say that, to their knowledge, this is the first open persona dataset at million-record scale that pairs Japan's official statistics with household finance and financial literacy data.

The dataset holds 9M personas, written in natural Japanese, organised as 1M records with 9 personas each. Each record has 50 fields: 9 persona fields, 11 persona attributes and 30 contextual fields based on official demographic and labor statistics. In total it runs to about 1.5B tokens, including about 750M persona tokens, with about 950k unique names and more than 1,600 occupation categories. The post says it holds 50% more personas than Nemotron-Personas-Japan and 50 fields instead of 22. Personas cover occupations, sports, arts, travel, cooking, finance, financial behavior and financial concerns, and the text attributes describe cultural background, skills, career goals, hobbies, household and financial background, attitudes toward money and financial planning.

It was built with NVIDIA's NeMo Data Designer, which the post describes as a system supporting complex Jinja templating, Pydantic validation, structured outputs, automated retries and multiple generation backends. Two models were used: NeMo SDG-PGMs (Apache-2.0), a probabilistic graphical model framework for statistically grounded generation, and gemma-4-31B-it (Apache-2.0) for the Japanese text. The design also took in insights and feedback from 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 (FDUA).

The team made several adjustments beyond the raw statistics. For education, it added finer granularity where national figures lump degree levels together. For occupation, it added categories such as business owners and specialized professions. For life stages, it modelled students, retirees and periods of unemployment, which rarely stand out in statistics. For household finances, it added household-level financial and real assets on top of individual income and spending propensity. For digital and financial literacy, it included internet and smartphone use alongside financial literacy.

On privacy, the post says every persona is fully synthetic, no real individuals are represented, and the dataset is designed to reduce privacy risk. Demographics and household finances follow official statistics and public survey data, and names follow real Japanese naming patterns. Personas are built from these distributions, not from records about real people.

The intended users are teams building financial AI in Japan, at banks, insurers and the startups and integrators that serve them. Suggested uses are customer-service agents, evaluating sales conversations and marketing initiatives, user simulations in NeMo User Sim, and red-teaming customer-facing AI. The post also lists multi-turn dialogue generation with personas as seeds, domain-specific assistants, fine-tuning on culturally grounded data, and bias testing across rural and urban divides, age groups, education levels and economic circumstances. The dataset is released under CC BY 4.0 and is available for commercial use.

A companion release, Nemotron-PII-Japan, launches the same day. It holds 100,000 synthetic Japanese business documents with labeled personal information, for training and testing detection and redaction models, and its documents are built on Nemotron-Personas-Japan. The post's suggested pairing: use the finance personas to simulate customers and the PII dataset to test how a pipeline protects their data.

Key facts

The dataset has 9M Japanese personas (1M records x 9 personas each), with 50 fields per record: 9 persona fields, 11 persona attributes and 30 contextual fields.
It totals about 1.5B tokens (about 750M persona tokens), with about 950k unique names and more than 1,600 occupation categories.
It is released under CC BY 4.0 and is available for commercial use; load it with load_dataset("nvidia/Nemotron-Personas-Finance-Japan").
Built with NeMo Data Designer, NeMo SDG-PGMs and gemma-4-31B-it (both Apache-2.0); design input came from Sompo Japan, Tokio Marine Holdings, Bank of Fukuoka, Mizuho Financial Group, Meiji Yasuda Life Insurance and FDUA.
Nemotron-PII-Japan, 100,000 synthetic Japanese business documents with labeled personal information, launches the same day.

Why it matters

Financial firms cannot freely use real customer data for AI work because of privacy regulation and compliance requirements. A synthetic population that carries household finances and financial literacy next to demographics gives them something to build and test against without touching customer records. The authors say that, to their knowledge, it is the first open persona dataset at million-record scale to pair Japan's official statistics with this kind of finance data. They also argue it reduces reliance on Western-centric datasets and supports Sovereign AI efforts. Those are the authors' claims.

Who it affects

The post names teams building financial AI in Japan first: banks and insurers, and the startups and integrators that serve them. It also says any Japanese-language project can use the finer population slices, since the dataset has 9M personas and 50 fields against 22 in Nemotron-Personas-Japan. Developers testing for bias across rural and urban areas, age groups, education levels and economic circumstances are another named audience.

How to use it

Download it with the Hugging Face datasets library: from datasets import load_dataset, then ds = load_dataset("nvidia/Nemotron-Personas-Finance-Japan"). The licence is CC BY 4.0, covering commercial and non-commercial use. The post suggests using personas as seeds for multi-turn dialogue generation, building customer-service agents, evaluating sales conversations and marketing initiatives, running simulations in NeMo User Sim, red-teaming customer-facing AI, fine-tuning on culturally grounded data, and bias testing. For privacy testing of a pipeline, it points to the companion Nemotron-PII-Japan dataset.

How solid is it

This is a first-party announcement from the releasers, so the claims are theirs. The personas are described as following official statistics and public survey data, and the dataset was built with named tooling and licences. The "first" claim is hedged by the authors themselves with "to our knowledge". No benchmark, evaluation or quality measurement of the dataset, or of models trained on it, is reported. The financial institutions listed provided insights and feedback that were incorporated into the design; the post does not say they use the dataset.

Risks and caveats

The data is fully synthetic and, in the authors' words, designed to reduce privacy risk; the post does not say the risk is eliminated. Claims that the data helps reduce biased training and the risk of model collapse are the authors' own and come without reported measurements. The personas follow aggregate statistics and survey data, and the team added categories and life stages beyond what the statistics show, so they are a modelled population, not a record of real people. The Japanese text was generated by gemma-4-31B-it.

“Every persona is fully synthetic. No real individuals are represented, and the dataset is designed to reduce privacy risk.”

— Elith and NVIDIA, Hugging Face blog post