What changed
NVIDIA today released the Nemotron PII Japan dataset, a collection of 100,000 synthetic Japanese business documents designed to help developers test privacy protection in AI applications. The dataset includes annotations identifying personally identifiable information (PII) and its types, allowing developers to examine missed detections and incorrect detection boundaries without using real customer information.
What this means in practice
The dataset is intended for developers working on AI applications that process Japanese business documents, such as financial institutions, healthcare providers, and government agencies. It provides a shared resource for developing and evaluating the safety of AI used in Japanese business workflows, including masking before sending documents to external LLMs, summarizing inquiry histories, and reviewing logs and reports before sharing.
The dataset contains documents across eight domains and 43 document types, including healthcare, finance, government, employment, education, legal, telecommunications and IT, and e-commerce.
Each document includes realistic business details such as loan purposes, requested loan amounts, dates of birth in Japanese era notation, addresses with building names and room numbers, and bank account details for repayment.
The PII and predetermined business values are managed as placeholders, inserted after text generation, and checked for consistency with the final text.
Limitations and availability
The dataset is available for commercial and noncommercial use under the CC BY 4.0 license. It stores 100,000 documents in a single data.parquet file, with both labeled field formats and natural prose formats. The dataset includes 56 PII entity types, such as names, addresses, email addresses, account numbers, and Japan's My Number IDs, with additional types like dates, dates and times, and times. The dataset also includes 810,703 positional and label annotations across 100,000 documents, or approximately 8.11 spans per document.
The dataset is designed to protect privacy by generating synthetic documents without collecting real customer or patient information. It includes quality checks for PII combinations, writing formats, and diversity of people and topics, in addition to document counts. The dataset is intended for developers to use in testing and evaluating AI applications that process Japanese business documents.
