Evaluating Privacy Protection and Practical AI Applications with 100,000 Synthetic Business Documents Across Eight Domains
Blog post from Hugging Face
NVIDIA released Nemotron PII Japan, a CC BY 4.0 dataset of 100,000 validated synthetic Japanese business documents designed to support privacy-preserving AI development without exposing real customer or patient data. Covering eight domains and 43 document types, including healthcare, finance, government, employment, education, legal, IT, and e-commerce, it contains 56 PII entity categories and more than 810,000 annotated PII spans, enabling evaluation of missed detections, excessive masking, and boundary errors while preserving useful business context. Documents are generated through a controlled pipeline that combines synthetic personas, domain catalogs, precomputed business values, placeholders, and post-generation consistency checks, including safeguards against unintended real information and labeling errors. The dataset supports workflows such as masking records before external LLM summarization, reviewing sensitive logs, and testing PII detection models, although its annotations are document-type-specific and require careful interpretation of precision metrics. NVIDIA also reports that fine-tuning a GLiNER model on the data substantially improved exact-span F1 on separate Japanese synthetic and out-of-distribution evaluation sets, while noting that synthetic benchmark results alone do not ensure production performance.
| Trend | Post Mentions | Total Month Mentions | Posts | Companies | MoM |
|---|---|---|---|---|---|
| LLM | 15 | 3,355 | 742 | 191 | -49% |
| AI Model Fine-tuning | 5 | 613 | 130 | 56 | -29% |
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.