The Pulse

Elith and NVIDIA Release 9M Synthetic Personas for Japan’s Financial AI

Elith and NVIDIA released Nemotron-Personas-Finance-Japan, a synthetic dataset of 9M Japanese personas for financial AI development. Its 50 fields cover demographic, household-finance and financial-literacy context, and the dataset is avail

AI.info Team · Written with AI · fact-checked · 2026-11-09

Nemotron-Personas-Finance-Japan contains 9M synthetic personas, organized as 1M records with nine personas each. Elith and NVIDIA released the dataset as a financial extension of the open-source Nemotron-Personas-Japan. NVIDIA says it is, to its knowledge, the first open persona dataset at million-record scale pairing Japan’s official statistics with household finance and financial-literacy data.

What each record holds

Each record has 50 fields: nine persona fields, 11 persona attributes and 30 contextual fields based on official demographic and labor statistics. The dataset contains ~1.5B tokens, including ~750M persona tokens, and ~950k unique names. Its 1600+ occupation categories reflect Japan’s workforce, with comprehensive coverage across demographic, regional, asset distribution and personality-trait axes. The diverse persona types include occupations, sports, arts, travel, cooking, finance, financial behavior and financial concerns.

How it was built

The dataset was built using NVIDIA’s NeMo Data Designer, which supports Jinja templating, Pydantic validation, structured outputs, automated retries and multiple generation backends. The generation pipeline used NeMo SDG-PGMs, an Apache-2.0 probabilistic graphical model framework, and gemma-4-31B-it (Apache-2.0) for Japanese text generation. The design includes finer educational categories than broad national statistics provide, expands occupation categories to include business owners and specialized professions, and includes life stages such as students, retirees and periods of unemployment. It also includes household-level financial and real assets, alongside individual income and spending propensity, plus internet and smartphone usage and financial literacy.

Financial teams can simulate customers and test models

The dataset is built first for teams creating financial AI in Japan, at banks and insurers and at the startups and integrators that serve them. Listed uses include building customer-service agents, evaluating sales conversations and marketing initiatives, running user simulations in NeMo User Sim, and red-teaming customer-facing AI. Developers can also use personas as seeds for multi-turn dialogue datasets, build financial assistants and test model performance across rural and urban divides, age groups, educational backgrounds and economic circumstances. The release holds 9M personas, 50% more than Nemotron-Personas-Japan, across 50 fields instead of 22.

The authors say the design drew on input from Japanese financial institutions and industry bodies: Sompo Japan, Tokio Marine Holdings, Bank of Fukuoka, Mizuho Financial Group, Meiji Yasuda Life Insurance and the Financial Data Utilization Association (FDUA). The aim, they write, is to let developers build and evaluate AI for local and industry needs without using real customer data.

Privacy and a companion dataset

Every persona is fully synthetic, represents no real individual and is designed to reduce privacy risk. Demographics and household finances follow official statistics and public survey data, while names follow real Japanese naming patterns. The dataset is released under CC BY 4.0 for commercial and non-commercial use. Nemotron-PII-Japan launches the same day and contains 100,000 synthetic Japanese business documents with labeled personal information for training and testing detection and redaction models. The finance personas can be used to simulate customers, and the PII documents can be used to test how a pipeline protects their data.