NVIDIA and Elith release nine million synthetic Japanese finance personas

NVIDIA and Elith released Nemotron-Personas-Finance-Japan, an openly licensed collection of nine million synthetic Japanese-language personas designed to help developers build and test financial AI without using real customer records.

The dataset contains one million records, each with nine persona fields, so the total is nine million personas rather than nine million separate records. NVIDIA and Elith describe each record as having 50 fields: the nine personas, 11 persona attributes and 30 contextual fields. They put the collection at about 1.5 billion tokens, including roughly 750 million persona tokens.

The dataset card labels the release CC BY 4.0 and says it can be used for commercial and noncommercial purposes. The license allows sharing and adaptation subject to attribution and its other conditions.

According to the creators, the generation pipeline first sampled attributes from public-statistics distributions with NVIDIA’s NeMo SDG-PGMs probabilistic graphical-model framework, which represents statistical relationships among variables. It then used NeMo Data Designer to generate personal context and persona text. Documented seed inputs included census data from Japan’s Statistics Bureau, Japanese first- and last-name data from Myoji-Yurai.net, and survey and statistical data from J-FLEC. The precise transformations were not independently audited.

The records pair demographic and geographic attributes with household composition, personal income, household financial assets, consumption propensity, financial literacy, held assets and securities mix. NVIDIA and Elith say every persona was generated from distributions, is not tied to a real person, and contains no financial-institution customer data or confidential information. Those claims were not independently testable from the published materials.

The creators position the data for dialogue generation, customer simulation, bias and fairness testing, red-teaming, and financial-agent evaluation. They also say the personas reflect Japanese demographic, geographic, household-finance, cultural and workforce distributions. No independent audit or peer-reviewed evaluation of that representativeness was identified in the research package.

The dataset card also documents limits to that fidelity. Its model makes simplifying independence assumptions, including treating occupation as independent of education once prefecture and sex are given, and it notes a lack of comprehensive census statistics on gender independent of biological sex. The creators describe distribution comparisons, automated validators and human review, but did not publish enough underlying results to reproduce their quality claims.

The NVIDIA-authored announcement appeared on November 9. The dataset card lists version 1.0 on November 10, when NVIDIA Japan announced the release at NVIDIA Agentic AI Day in Tokyo.