NVIDIA releases Nemotron-Personas-Finance-Japan, 9M-record open synthetic dataset for financial AI

NVIDIA and Japanese synthetic data specialist Elith have published Nemotron-Personas-Finance-Japan, a million-record-scale open dataset designed to train and evaluate financial-domain AI without exposing real customer information. The dataset contains 9 million total personas (1 million base records multiplied across 9 persona variants each), with 50 fields per record covering demographics, labor statistics, household finances, and financial literacy indicators. It was constructed using NVIDIA's NeMo Data Designer pipeline, the Apache-2.0 licensed NeMo SDG-PGMs probabilistic graphical model framework, and Google's Apache-2.0 licensed Gemma 4 31B instruction-tuned model for Japanese natural language generation.

The release addresses a structural barrier in Japanese financial AI development: strict privacy regulations and compliance requirements that limit use of real customer data for training. Input from six major domestic institutions including Mizuho Financial Group, Sompo Japan, and the Financial Data Utilization Association shaped the dataset's design. The CC BY 4.0 license permits commercial application, and integration with NVIDIA's NeMo User Sim framework enables immediate deployment for customer simulation and conversational agent evaluation.

Dataset architecture and generation pipeline

• Scale and structure: 9 million personas derived from 1 million base records with 9 variants each, totaling approximately 1.5 billion tokens including 750 million persona-specific tokens, with 50 structured fields per record.
• Statistical grounding: Built on NeMo SDG-PGMs, an Apache-2.0 licensed probabilistic graphical model framework that ensures generated data aligns with official Japanese government demographic and labor distributions rather than LLM hallucination.
• Generation infrastructure: NeMo Data Designer provides Jinja templating, Pydantic validation, structured output formatting, automated retry logic, and multi-backend support for industrial-scale synthetic data pipelines.
• Language model backend: Japanese text generation uses Google's Gemma 4 31B instruction-tuned model, also Apache-2.0 licensed, rather than NVIDIA's own proprietary models.
• Privacy architecture: Designed as "private by design" with no real customer data incorporated; synthetic construction avoids re-identification risks while preserving statistical properties needed for financial AI training.

Why it matters

The release demonstrates a viable pattern for regulated-domain open data: combining open-source probabilistic methods with open-weights language models to produce commercially usable synthetic datasets that preserve demographic fidelity without privacy exposure. For Japanese developers and financial institutions, it removes a significant data access barrier that has slowed domestic AI adoption relative to less regulated sectors, while the CC BY 4.0 license and Apache-2.0 tooling stack ensure no vendor-specific inference dependency is required for dataset generation or downstream use.