GLiNER-PII is a non-generative model created by NVIDIA for identifying and classifying Personally Identifiable Information (PII) and Protected Health Information (PHI) in both structured and unstructured text. Rather than producing free-form prose, it returns span-level entity annotations together with confidence scores, drawing on a taxonomy of more than 55 sensitive-data categories. Architecturally, it is inspired by the Gretel GLiNER PII/PHI work and is built on the GLiNER large-v2.1 base, inheriting that encoder-only span-classification design rather than a generative language-modeling setup. This lineage makes it well suited for token-aware entity tagging at scale instead of open-ended content generation.
The model is positioned for sensitive-information detection and redaction inside regulated and enterprise pipelines. Documented scenarios include redacting PHI from clinical notes and medical reports, identifying account numbers and Social Security numbers within banking and insurance files, protecting client information in legal contracts and filings, and scanning corporate documents, emails, and data stores for data-governance and compliance workflows such as GDPR, HIPAA, and CCPA. Use is governed by the NVIDIA Open Model License Agreement, and the model is designated for global commercial and non-commercial deployment, giving teams a uniform detector that can be embedded into content moderation, cybersecurity, and privacy-engineering stacks where consistent span-level labelling matters more than generative reasoning.
