NVIDIA and Elith release a synthetic Japanese finance persona dataset

The dataset extends the open Nemotron-Personas-Japan collection into finance. It contains 1 million records with 9 personas each, for 9 million personas in total, and 50 fields per record covering demographics, household finances and financial literacy. The natural Japanese text totals about 1.5 billion tokens.

NVIDIA built the dataset with its NeMo Data Designer and NeMo SDG-PGMs frameworks, and used gemma-4-31B-it for Japanese text generation. The personas are aligned with Japanese public demographic and labor statistics and are described as fully synthetic, not modeled on real individuals. Sompo Japan, Tokio Marine Holdings, Fukuoka Bank, Mizuho Financial Group, Meiji Yasuda Life and FDUA provided feedback during development.

The dataset is designed to work with the NeMo User Sim user simulation framework. A companion dataset, Nemotron-PII-Japan, was released the same day, containing 100,000 labeled synthetic Japanese business documents for training and testing personal information detection and masking models. The dataset loads with the Hugging Face datasets library using nvidia/Nemotron-Personas-Finance-Japan.

Teams building Japanese finance agents can test customer service flows, sales scripts and red-teaming against a large set of personas without collecting customer data. Developers can also use the persona fields as seeds for multi-turn conversation generation, or check whether an agent performs consistently across urban and rural, age and income groups. Because the personas are synthetic, results should still be validated against real Japanese customer interactions before deployment.