Nvidia and Elith Release 9 Million Synthetic Japanese Personas for Financial AI

The dataset is a finance-focused extension of the earlier Nemotron-Personas-Japan. The authors wrote that, to their knowledge, it is the first open persona dataset at million-record scale to combine Japan's official statistics with household finance and financial literacy data. Several large Japanese financial groups gave input on the design, including Sompo Japan, Tokio Marine Holdings, Bank of Fukuoka, Mizuho Financial Group, Meiji Yasuda Life Insurance and the Financial Data Utilization Association.

The release targets a practical problem for regulated industries. Financial firms want to train and test customer-facing AI agents, but privacy rules and compliance requirements limit how they can use real customer data. In their blog post, the authors argue that synthetic data can avoid those risks if it reflects the demographics, household structures and assets of the real population. Every persona is fully synthetic and represents no real person, they wrote. The figures are drawn from distributions in official statistics and public surveys, not from records about individuals.

The dataset has 1 million records with nine personas each, written in natural Japanese. It contains about 1.5 billion tokens, about 750 million of them persona text, along with roughly 950,000 unique names and more than 1,600 occupation categories. Each record has 50 fields, compared with 22 in the earlier Japanese release. These include household-level financial and real assets, income and spending habits, internet and smartphone use, financial literacy, and written descriptions of attitudes toward money and financial planning.

The developers said they went beyond the raw government data in places. They added finer gradations of education, extra occupation types such as business owners and specialised professionals, and life stages that national statistics tend to underrepresent, including students, retirees and people between jobs.

The data was generated with Nvidia's NeMo Data Designer and NeMo SDG-PGMs, a probabilistic graphical model framework for grounding synthetic records in statistical distributions. Google's open gemma-4-31B-it model wrote the Japanese text. The personas are designed to work with NeMo User Sim, Nvidia's tool for simulating users.

The suggested uses include building customer-service agents, simulating customers to test sales conversations and marketing campaigns, generating multi-turn dialogue for training, and red-teaming customer-facing AI before it reaches real clients. The blog post gives scale figures but no benchmark results showing how models trained or evaluated on the data perform.

The dataset joins a growing set of Nvidia releases aimed at Japanese developers, including Nemotron-PII-Japan, a collection of 100,000 synthetic business documents labelled for personal information. It can be downloaded from Nvidia's page on the Hugging Face hub.