NVIDIA releases two synthetic Japanese datasets, Nemotron-Personas-Finance-Japan and Nemotron-PII-Japan

NVIDIA has introduced two synthetic datasets to support Japanese-language AI development: Nemotron-Personas-Finance-Japan and Nemotron-PII-Japan. The company explained them on November 10, 2026 at NVIDIA Agentic AI Day Japan 2026, a technical event on AI agents held in Tokyo. Both datasets are published on Hugging Face. The source is a Japanese-language report by MONOist (ITmedia).

Nemotron-Personas-Finance-Japan

The dataset extends the existing Nemotron-Personas-Japan into the financial domain. Beyond demographic attributes, household composition, assets, income form and holdings, the personas include financial context such as attitudes toward investment, life plans and concerns about money. NVIDIA says the personas are intended for financial AI development and user simulation.

The dataset was developed jointly with Elith, an AI startup. To validate practicality in financial use cases, NVIDIA says it received cooperation from several financial and insurance organizations, including Mizuho Financial Group.

Nemotron-PII-Japan

Nemotron-PII-Japan is a synthetic dataset built from Japanese documents that contain personally identifiable information (PII). Key specifications from the report:

• Domains: 8, covering healthcare, finance, government, employment, education, legal, IT/security and e-commerce.
• Document types: 43, including diagnostic documents and address change notices.
• PII levels: 56.
• Annotations: each document marks where PII appears and which category it belongs to.

Because no real individuals' data is used, the dataset provides text and annotations together at the level of individual documents. NVIDIA positions it for additional training, evaluation and improvement of models that detect and mask PII in Japanese.

Context within the Nemotron family

Nemotron is described as an open multimodal AI model family developed by NVIDIA for building systems with AI agents. It also includes datasets, tools, libraries and reference designs called Blueprints. According to the report, Nemotron 1 was announced in 2024 by adding reasoning ability to the open-source LLM Llama. Nemotron 2 followed in 2025, trained from scratch, and Nemotron 3 in 2026 is described as capable of executing advanced AI agent tasks.

The two new datasets follow earlier Japan-focused work led by a senior AI solutions architect at NVIDIA. That work includes Nemotron-Personas-Japan, an open dataset of personas reflecting Japan's demographics, geographic distribution and personality traits. It also includes Nemotron-Nano-9B-v2-Japanese, a version of the small language model Nemotron-Nano-9B-v2 with advanced Japanese comprehension and AI agent functions.

According to the report, the Japanese dataset work runs in parallel with Nemotron development centered in the US. The architect said that when issues arise, the team can consult Nemotron's development group and, in some cases, NVIDIA's internal staff. He described close coordination between Japan and the US as a strength of NVIDIA's approach.

What this means

NVIDIA is extending Nemotron beyond model weights into domain-specific Japanese data. The PII dataset's document-level annotations (PII location and category) make it directly usable for training and evaluating detection and masking models without handling real personal data. The finance personas dataset gives developers of financial AI a way to run user simulations using Japanese-specific financial context. Both are open on Hugging Face, so teams can test them directly. The report does not name the dataset licenses, so users should check the Hugging Face pages.