Content Deep Dive

Nemotron PII Japan: プライバシー保護に向けた日本語PIIオープンデータ

Blog post from Hugging Face

Post Details
Company
Date Published
Author
Masaya Ogushi, Atsunori Fujita, Vincent Gong, Kotaro Yamamoto, Colleen Ruan, Yoshi Suhara, Will Jennings, and Andre Manoel
Word Count
462
Company Posts That Month
60
Language
-
Hacker News Points
-
Post removed?
No
Summary

NVIDIA released Nemotron PII Japan, an open CC BY 4.0 dataset of 100,000 privacy-preserving synthetic Japanese business documents designed to evaluate PII detection, masking, and context-preserving AI workflows without using real customer data. Covering 43 document types across medical, finance, government, employment, education, legal, telecommunications/IT, and e-commerce domains, it contains about 34.7 million tokens and 810,703 annotated PII spans across 56 entity types, including names, addresses, account numbers, dates, and organization categories. The dataset was generated with NeMo Data Designer using synthetic personas, curated domain catalogs, precomputed calculation plans, and placeholders that insert validated PII and business values after text generation, helping ensure consistency between annotations, document content, and numerical details such as loan repayment calculations. It supports testing applications such as pre-LLM redaction, inquiry summarization, and secure sharing of logs or reports, while cautioning that annotations are governed by document-specific rules rather than exhaustive labeling and that synthetic-data performance does not guarantee production performance. Fine-tuning the GLiNER model on the dataset substantially improved external Japanese PII evaluation results, increasing micro F1 from 13.41% to 73.80% on Sumi test data and from 8.13% to 65.30% on out-of-distribution Sumi data.

Trends Found in this Post

No tracked trend matches for this post yet.

Use This Data

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.