Elith and Nvidia have released Nemotron-Personas-Finance-Japan, an open synthetic dataset for building and evaluating financial AI in Japanese contexts. It contains nine million personas derived from one million records, with nine persona variants per record, and is available for commercial and noncommercial use under CC BY 4.0.
Each record has 50 fields covering demographics, employment, household finances, assets, financial literacy and behavior. The collection totals roughly 1.5 billion tokens and includes more than 1,600 occupation categories. Official government statistics and public surveys shape the distributions, while Gemma 4 31B generated Japanese text through Nvidia’s synthetic-data tools.
The creators say no real individuals are represented. Banks, insurers and developers can use the personas to simulate customers, generate conversations, test sales or service agents and compare performance across ages, regions, education levels and economic circumstances. Several Japanese financial institutions and an industry association provided feedback on the design.
Synthetic records reduce the need to expose customer data, but they do not eliminate bias. Added categories, modeled life stages and generated language reflect design choices as well as public statistics. Teams should therefore validate whether results transfer to real users before using the dataset for consequential financial decisions. A companion release provides 100,000 synthetic business documents for testing detection and redaction of personal information.