Nemotron-Personas-Japan: Synthesized Data for Sovereign AI

A compound AI approach to Japanese personas grounded in real-world distributions

Open Data for Japan's AI Future

Building AI that truly understands Japanese culture has been nearly impossible without authentic, diverse training data. Today, we're changing that. Nemotron-Personas-Japan is the first open synthetic dataset that captures Japan's demographic, geographic, and cultural spectrum. Licensed under CC BY 4.0, this dataset provides a privacy-preserving, regulation-ready foundation for AI systems that reflect Japanese society without relying on sensitive personal data.

Created with NeMo Data Designer, NVIDIA’s enterprise-grade system for synthetic data generation, Nemotron-Personas-Japan builds on the success of the widely used US Personas dataset. Together, these releases mark the beginning of a global collection of synthetic persona datasets and playbooks that support sovereign AI development across countries and regions.

The dataset is designed to work seamlessly with Nemotron models and other open-source LLMs, making it easy to fine-tune for Japanese AI applications - from enterprise chatbots to domain-specific copilots.

What’s in the Dataset?

• 6M personas total (1M records x 6 personas each), written in natural Japanese
• 22 fields per record: 6 persona fields and 16 contextual fields grounded in official demographic and labor statistics
• ~1.4B tokens total, including ~850M persona tokens
• ~950k unique names – unprecedented diversity in synthetic data generation
• 1500+ occupation categories reflecting Japan's workforce
• Comprehensive coverage of demographic, geographic, and personality trait axes
• Variety of persona types: professional, sports, arts, travel, culinary
• Natural language persona attributes: cultural background, skills & expertise, goals & ambitions, hobbies & interests
• Licensed under CC BY 4.0 for full commercial and non-commercial use

How We Built It

Data Generation Pipeline

Built with NeMo Data Designer, NVIDIA’s microservice for synthetic data generation. This compound AI system enables generation with complex Jinja templating, Pydantic validation, structured outputs, automated retries, and supports multiple generation backends – the necessary tooling to scale a synthetic dataset of this size. We also leveraged the following models:

1. Probabilistic Graphical Model (Apache-2.0) for statistical grounding
2. GPT-OSS-120B (Apache-2.0) for narrative generation in Japanese

Enhanced Cultural Context

Nemotron-Personas-Japan was designed to align with Japan’s official demographic and labor statistics, while extending them into areas important for AI training. In practice, this meant:

• Education: Where degree levels are grouped in national statistics, we introduced finer distinctions so models can reflect different educational pathways.
• Occupations: We incorporated additional categories (such as business owners and specialized trades) to broaden the occupational spectrum used in training.
• Life Stages: We included student, retirement, and unemployment status information which are important for realistic personas.
• Cultural Traits: To ensure authenticity, we included Japanese social and cultural characteristics that help AI systems better reflect local norms.
• Digital Divide: We accounted for different levels of digital literacy across age groups to reflect real-world technology usage patterns in Japan.

Private By Design

This dataset does not contain any personally identifiable information (PII). While we use real-world distributions of ages, names, and occupations from official public sources, nothing is ever tied to any real person, living or deceased. Every persona is fully synthetic, allowing you to train on authentic real-world cultural patterns without compromising personal privacy.

Who This Data Is For

Nemotron-Personas-Japan is designed first and foremost for Japanese model builders developing sovereign AI systems. Most training data used by LLM builders today is in English, leaving local developers in Japan, India, and other regions struggling to source high-quality data in their native languages.

Our Nemotron-Personas effort directly addresses this challenge by helping model builders generate diverse, complex data in their local language while capturing crucial region-specific nuances. We ground our datasets in local context—census data, naming conventions, cultural patterns—and produce everything in the native language.

That said, global developers should absolutely leverage this data if they want their models to achieve better adoption in Japan and understand Japanese cultural contexts.

Practical AI Applications

Here’s how you can put these synthetic personas to work today:

• Multi-Turn Conversation – Use personas as seeds to create authentic dialogue datasets.
• Domain-specific Training – Create training datasets for building culturally aware AI assistants
• Bias-Testing & Fairness – Evaluate how your models and agentic systems perform across rural vs urban populations, different age groups, or various education levels - ensuring your AI works fairly for all segments of Japanese society.

Why It Matters

Open-source AI development has long struggled with access to diverse, high-quality training data that reflects real-world populations. Proprietary datasets dominate enterprise AI, creating barriers for researchers, startups, and developers in underrepresented regions.

• Data Diversity: Prevents narrow training and model collapse by reflecting Japan’s full population spectrum.
• Cultural Authenticity: Reduces reliance on Western-centric datasets, and supports the development of Sovereign AI systems.
• Privacy & Compliance: Meets Japan’s PIPA requirements and future AI governance standards.

By releasing Nemotron-Personas-Japan under CC BY 4.0, we're democratizing access to enterprise-grade synthetic data, enabling anyone to build culturally authentic AI systems without the typical barriers of cost, privacy concerns, or geographic limitations.

Start Building with Nemotron-Personas-Japan

Ready to create AI that truly understands Japanese culture and language?

To start experimenting today:

from datasets import load_dataset

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

For production applications:

• Use personas as seeds for conversation generation
• Fine-tune models on culturally-grounded data
• Build personalization engines that reflect Japan's full demographic spectrum
• Develop domain-specific copilots with authentic Japanese context

Whether you're a Japanese model builder developing Sovereign AI or a global developer seeking better regional adoption, Nemotron-Personas-Japan dataset provides the authentic, privacy-safe foundation your applications need.