Nvidia releases a new synthetic dataset for Japanese financial AI

Nvidia and Elith have published Nemotron-Personas-Finance-Japan, a collection of 9 million synthetic personas grounded in official Japanese statistics. The release addresses privacy constraints in the financial sector by offering a million-record scale dataset that mirrors demographics, household structures, and asset statuses without using real customer data.

How it works

The team used NeMo Data Designer and NeMo SDG-PGMs to generate the data. These tools apply probabilistic graphical models to ensure statistical accuracy. Development included feedback from Sompo Japan, Tokio Marine Holdings, Bank of Fukuoka, Mizuho Financial Group, Meiji Yasuda Life Insurance, and the Financial Data Utilization Association.

Released under the CC BY 4.0 license, the dataset supports commercial use. It integrates with the NeMo User Sim framework to run instant simulations based on diverse distributions.

Dataset specifications

The collection contains the following elements:

• 9M total personas, structured as 1M records with 9 personas each
• 50 fields per record, split into 9 persona fields, 11 persona attributes, and 30 contextual fields
• Approximately 1.5 billion tokens, including roughly 750 million persona tokens
• Nearly 950,000 unique names
• Over 1,600 occupation categories reflecting the Japanese workforce
• Coverage across demographics, regions, asset distribution, and personality traits
• Categories covering occupations, sports, arts, travel, cooking, finance, financial behaviour, and financial concerns
• Natural language attributes for cultural background, skills, career goals, hobbies, household finances, and attitudes toward money

Construction details

The generation pipeline relies on NeMo Data Designer. This system handles complex Jinja templating, Pydantic validation, structured outputs, and automated retries. The process used two specific models:

• NeMo SDG-PGMs (Apache-2.0): A probabilistic graphical model framework for statistically grounded data generation
• gemma-4-31B-it (Apache-2.0): Used for Japanese text generation

Enhanced cultural context

The dataset aligns with official Japanese demographic and labor statistics while adding specific layers for AI training:

• Education: Finer granularity replaces broad degree aggregates to reflect diverse pathways
• Occupation: Additional categories include business owners and specialized professions
• Life Stages: Scenarios for students, retirees, and unemployment periods are modelled to improve realism
• Household Finances: Individual income and spending propensity are combined with household-level assets
• Digital & Financial Literacy: Internet and smartphone usage data capture consumer profiles more precisely

Privacy approach

Every persona is fully synthetic. No real individuals are represented. Demographics and finances follow public survey data, while names follow real Japanese patterns. The personas are built from these distributions rather than records about actual people.

Target audience

The primary users are teams building financial AI for Japanese banks, insurers, and their startup partners. Each persona includes household finances and financial literacy alongside demographic data. Applications include customer-service agents, sales conversation evaluation, marketing initiative testing, and red-teaming customer-facing AI.

The release also serves anyone building for Japanese users. This version holds 9 million personas, which is 50% more than the previous Nemotron-Personas-Japan dataset. It provides 50 fields instead of 22, offering finer population slices for training and testing.

Practical applications

Developers can use the synthetic personas for the following tasks:

• Multi-turn conversation: Using personas as seeds to generate human-like dialogue datasets
• Domain-specific AI development: Creating assistants tailored to realistic financial scenarios
• Bias testing and fairness: Evaluating model performance across rural and urban divides, age groups, educational backgrounds, and economic circumstances

Why it matters

Access to diverse training data that reflects real populations remains difficult. Enterprise AI has often relied on private data, creating barriers for researchers and startups. This release helps reduce biased training and the risk of model collapse.

The cultural authenticity reduces reliance on Western-centric datasets and supports Sovereign AI systems. Privacy and governance align with Japan’s innovation-first approach because every persona is synthetic, removing the need for customer data to build or test.

Nvidia and Elith are releasing the data under CC BY 4.0 to lower costs and privacy risks for building culturally grounded systems.

Related release

Nemotron-PII-Japan launches on the same day. It contains 100,000 synthetic Japanese business documents with labeled personal information for training detection and redaction models. The documents are built on Nemotron-Personas-Japan. Teams can use Nemotron-Personas-Finance-Japan to simulate customers and Nemotron-PII-Japan to test data protection pipelines.

Getting started

You can download the dataset using the following command:

from datasets import load_dataset
ds = load_dataset("nvidia/Nemotron-Personas-Finance-Japan")

Use cases for production applications include using personas as seeds for dialogue generation, fine-tuning models with culturally grounded data, building personalized engines that reflect Japan’s demographic makeup, and developing domain-specific agents equipped with Japanese context.

What it means

Japanese developers now have a public resource to test financial AI without privacy violations. The dataset allows teams to evaluate how systems handle different economic backgrounds and financial literacy levels. This supports the creation of agents that understand local nuances without risking customer data.