Nemotron-PII is a synthetic dataset released by NVIDIA in 2025 for privacy-preserving AI, designed to train and evaluate the detection and de-identification of personally identifiable information (PII) and protected health information (PHI) in text.

The dataset contains 100,000 English records spanning over 50 industry domains, generated using synthetic personas based on U.S. Census data to ensure demographic realism and contextual consistency. It includes span-level annotations for more than 55 PII/PHI categories, covering both structured (e.g., tables, invoices) and unstructured (e.g., emails, free text) documents, with explicit labeling of U.S. or international localization conventions.

Dataset Composition

• uid: String/UUID
• domain: String, representing the industry domain
• document_type: String, document type
• document_description: String, document description
• document_format: String, structured or unstructured
• locale: String, us or intl
• text: UTF-8 string, document content
• spans: List, containing entity boundaries and labels
• text_tagged: String, inline-tagged text

Citation

@dataset{nemotron-pii,
  author = {Amy Steier and Andre Manoel and Alexa Haushalter and Maarten Van Segbroeck},
  title = {Nemotron-PII: Synthesized Data for Privacy-Preserving AI},
  year = {2025},
  publisher = {NVIDIA},
  url = {https://huggingface.co/datasets/nvidia/Nemotron-PII}
}