Data

NVIDIA Releases Nemotron Personas Finance Japan and Nemotron PII Japan, Two Synthetic Datasets for Japanese AI Agents

· 6 min read · Data

NVIDIA Releases Nemotron Personas Finance Japan and Nemotron PII Japan, Two Synthetic Datasets for Japanese AI Agents

NVIDIA has released two synthetic datasets for developers building Japanese-language AI agents, the company’s Japan blog announced on November 10 during its NVIDIA Agentic AI Day developer event in Tokyo. Nemotron Personas Finance Japan extends an earlier collection of synthetic Japanese personas into the financial domain, while Nemotron PII Japan is designed to help train models that find and mask personal information in Japanese documents. Both are available on Hugging Face as part of the open Nemotron family.

The finance set builds on Nemotron Personas Japan, released in 2025, which generated 6 million synthetic personas from 1 million records based on public population and labour statistics. The new version keeps those base attributes and adds income, assets, spending, saving and investment behaviour, designed with input from Japanese financial institutions. It raises the persona count to 9 million and total tokens from about 1.4 billion to about 1.5 billion. NVIDIA says every person in it is artificially generated and cannot be linked to a real individual. Elith, an AI startup that emerged from the Matsuo Institute linked to the University of Tokyo, handled implementation and large-scale generation.

What it means in practice

NVIDIA Releases Nemotron Personas Finance Japan and Nemotron PII Japan, Two Synthetic Datasets for Japanese AI Agents — contextual photo

Nemotron PII Japan contains about 100,000 fully synthetic documents across healthcare, finance, government, employment, education, law, IT security and e-commerce, covering 43 document types such as medical certificates, loan applications and change-of-address forms. Each document is annotated to show where names, addresses, email addresses and identification numbers appear and which category they belong to. NVIDIA stresses that it is releasing a dataset, not a finished Japanese PII model, so companies can train and test masking tools without moving real customer data. Financial firms could use the persona set to simulate multi-turn conversations, market research and sales training.

The releases add to a growing pool of open resources for Japanese AI, following Preferred Networks’ open-weight PLaMo 3 Translate and finance-sector agent projects such as NEOBANK ai BIZ from Docomo SMTB Net Bank. What to watch is whether Japanese banks and insurers adopt the persona data to test customer-facing agents, and whether domestic developers publish PII-masking models trained on the new documents.

← Back to AI4Japan