Banks and insurers that want to build AI assistants face a basic obstacle: the data they would need to test them on is exactly the customer data privacy rules forbid them to use freely. Nvidia and the Japanese firm Elith have released a dataset aimed at that gap in Japan, a large set of invented but statistically grounded customers that anyone can use, including commercially, under a CC BY 4.0 licence.
Nemotron-Personas-Finance-Japan holds 1 million records, each with 9 personas, for 9 million in total, written in natural Japanese. Each record has 50 fields: persona descriptions, attributes such as attitudes to money and financial planning, and 30 context fields drawn from official demographic and labour statistics. The set totals about 1.5 billion tokens, with roughly 950,000 unique names and more than 1,600 occupation categories. It extends the earlier Nemotron-Personas-Japan, which had 22 fields per record, and adds household assets, financial literacy and internet and smartphone use.
The build pipeline is open about its parts. A probabilistic graphical model, NeMo SDG-PGMs, samples attributes so they follow government distributions, and Google's open gemma-4-31B-it model writes the Japanese text, all orchestrated through Nvidia's NeMo Data Designer. The team says major Japanese institutions, including Sompo Japan, Tokio Marine, Mizuho and Meiji Yasuda Life, gave feedback on the design. The suggested uses are seeding customer-service dialogues, simulating users in NeMo User Sim, and checking whether a model treats rural and urban, young and old, or richer and poorer customers consistently. A sister release, Nemotron-PII-Japan, offers 100,000 synthetic business documents with labelled personal information for testing redaction.
This is part of a pattern. Nvidia has been releasing country-specific persona sets that feed its own tooling, and synthetic users are becoming a standard way to test chat agents before real people see them. For Japanese-language work in general, a large, permissively licensed set with fine demographic slices is useful even outside finance.
The limits are built in. Personas drawn from official statistics reproduce the averages and spreads of those statistics; they will miss the unusual customers and messy situations that cause real support failures. The text is written by a single 31B model, which brings its own habits of phrasing. The announcement offers no evaluation showing that agents tested on these personas do better with real customers, so the gain for any given bank remains to be shown.