Nemotron-Personas-Finance-Japan: A Synthetic Dataset for Japan’s Demographics and Financial Context

Nemotron-Personas-Finance-Japan: A Synthetic Dataset for Japan’s Demographics and Financial Context

Published November 9, 2026

Open Data for Financial AI Development

Financial organizations face substantial limits when using real-world data because of privacy regulations, compliance requirements, and the specialized expertise needed to handle financial information. Developing agentic AI for finance therefore requires synthetic data that avoids the risks of real customer records while still representing demographics, household structures, asset ownership, and varied financial situations.

Elith and NVIDIA released Nemotron-Personas-Finance-Japan, a financial extension of the open-source Nemotron-Personas-Japan dataset. The release is described as the first open persona dataset at million-record scale to combine Japan’s official statistics with household finance and financial literacy data.

The dataset was generated with NeMo SDG-PGMs, a probabilistic graphical model framework for synthetic data generation, and NeMo Data Designer, an enterprise synthetic data generation system. Its design also incorporated feedback from Japanese financial institutions and industry associations, including Sompo Japan, Tokio Marine Holdings, Bank of Fukuoka, Mizuho Financial Group, Meiji Yasuda Life Insurance, and the Financial Data Utilization Association (FDUA).

Released under the CC BY 4.0 license, the dataset is available for commercial and non-commercial use. It supports conversational agents, customer simulation, model evaluation, and other financial AI development tasks. The Nemotron-Personas collection, including Nemotron-Personas-Finance-Japan, is designed to work with the NeMo User Sim user-simulation framework, allowing simulations based on distributions intended to reflect real-world populations.

Dataset Contents

Nemotron-Personas-Finance-Japan includes:

• 9 million personas in total, consisting of 1 million records with nine personas each, written in natural Japanese
• 50 fields per record, including nine persona fields, 11 persona attributes, and 30 contextual fields based on official demographic and labor statistics
• Approximately 1.5 billion tokens, including about 750 million persona tokens
• Approximately 950,000 unique names
• More than 1,600 occupation categories representing Japan’s workforce
• Coverage of demographic, regional, asset-distribution, and personality-trait dimensions
• Persona types covering occupations, sports, arts, travel, cooking, finance, financial behavior, and financial concerns
• Natural-language attributes describing cultural background, skills and expertise, career goals and aspirations, hobbies and interests, household and parental background, financial background, attitudes toward money, and financial planning
• Commercial and non-commercial use under the CC BY 4.0 license

How the Dataset Was Built

Data Generation Pipeline

The dataset was created with NVIDIA’s NeMo Data Designer synthetic data generation framework. The compound AI system supports complex Jinja templating, Pydantic validation, structured outputs, automated retries, and multiple generation backends, which are used to manage large-scale synthetic data production.

The generation process also used:

• NeMo SDG-PGMs, licensed under Apache-2.0, for statistically grounded data generation with probabilistic graphical models
• gemma-4-31B-it, licensed under Apache-2.0, for Japanese text generation

Cultural and Financial Context

The dataset was structured around Japan’s official demographic and labor statistics, with additional detail intended to support AI training and financial applications:

• Education: Finer-grained education levels were introduced where national statistics use broader degree categories.
• Occupation: Additional categories, including business owners and specialized professions, broaden the occupational range.
• Life stages: Students, retirees, and periods of unemployment are represented alongside more common working-life scenarios.
• Household finances: The dataset includes household-level financial and real assets in addition to individual income and spending propensity.
• Digital and financial literacy: Internet use, smartphone use, and financial literacy help represent consumer profiles relevant to financial services.

Synthetic Data and Privacy

Every persona is synthetic, and the dataset does not represent specific real individuals. Demographic and household-finance distributions are based on official statistics and public survey data, while names follow real Japanese naming patterns. The personas are generated from these distributions rather than from records about actual people.

Intended Users and Applications

The dataset is intended for teams developing financial AI in Japan, including banks, insurers, startups, and system integrators. Because each persona combines demographics with household finances and financial literacy, teams can use the data to:

• Build customer-service agents
• Evaluate sales conversations and marketing initiatives
• Run user simulations with NeMo User Sim
• Red-team customer-facing AI systems
• Develop applications for Japanese-language users

Compared with Nemotron-Personas-Japan, the dataset contains 9 million personas, 50 percent more, and expands the schema from 22 to 50 fields. This provides more detailed population segments for training and testing Japanese-language systems.

Practical AI Applications

The synthetic personas can support several development activities:

• Multi-turn conversation: Use personas as seeds for generating human-like dialogue datasets.
• Domain-specific AI: Create training and evaluation data for assistants operating in realistic financial scenarios.
• Bias and fairness testing: Assess model and agent performance across rural and urban areas, age groups, education levels, and economic circumstances.

Why the Dataset Matters

Access to diverse, high-quality data that reflects real populations remains a challenge in AI development. Enterprise projects have often depended on private data, creating barriers for researchers, startups, and developers working in regions with limited data availability.

Nemotron-Personas-Finance-Japan addresses these constraints by offering:

• Population diversity: Coverage of Japan’s population across multiple demographic and financial dimensions
• Cultural context: Japanese-language personas grounded in local demographic, occupational, household, and financial conditions
• Privacy and governance: Synthetic records that can be used without customer data for development and testing

The dataset is released under CC BY 4.0 and is available for commercial use.

Relationship to Nemotron-PII-Japan

Nemotron-PII-Japan was released on the same day. It contains 100,000 synthetic Japanese business documents with labeled personal information for training and testing personal-information detection and redaction models. Its documents are built on Nemotron-Personas-Japan.

The two datasets support different stages of testing: Nemotron-Personas-Finance-Japan can simulate customers, while Nemotron-PII-Japan can be used to evaluate how data-protection pipelines handle personal information.

Loading the Dataset

The dataset can be loaded with the Hugging Face datasets library:

from datasets import load_dataset

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

Potential development uses include generating dialogue data, fine-tuning models with culturally grounded examples, building personalization systems that reflect Japan’s demographic composition, and developing financial agents with Japanese context.