Nemotron-Personas-Finance-Japan: A Synthetic Dataset Capturing Japan's Demographics and Financial Context

NVIDIA and Japanese AI company Elith have released Nemotron-Personas-Finance-Japan, an openly licensed synthetic dataset designed to help developers build and test financial AI without touching real customer data. Announced on Hugging Face on November 9, 2026, it extends the earlier Nemotron-Personas-Japan release, and the authors describe it as, to their knowledge, the first million-record open persona dataset combining Japan's official statistics with household finance and financial literacy information.

The dataset contains 1 million records, each with nine personas, for 9 million personas in natural Japanese. Records carry 50 fields, up from 22 in the predecessor, covering demographics, occupations across more than 1,600 categories, household assets, attitudes toward money, and digital and financial literacy. It totals roughly 1.5 billion tokens and about 950,000 unique names. NVIDIA says it was generated with its NeMo Data Designer system and NeMo SDG-PGMs probabilistic modeling framework, using Google's gemma-4-31B-it for Japanese text, and was shaped by feedback from institutions including Mizuho Financial Group, Tokio Marine Holdings, Sompo Japan, Bank of Fukuoka, Meiji Yasuda Life Insurance and the Financial Data Utilization Association.

The authors stress that every persona is fully synthetic, built from statistical distributions rather than records about real people. Suggested uses include customer-service agents, sales-conversation evaluation, user simulation through NeMo User Sim, red-teaming and fairness testing across age, region and income. It is released under CC BY 4.0, permitting commercial use.

A companion dataset, Nemotron-PII-Japan, launched the same day with 100,000 synthetic Japanese business documents labeled for personal information, aimed at training detection and redaction models.