pii-proxy

A GLiNER model fine-tuned for PII detection. GLiNER does zero-shot NER over arbitrary labels — you pass the entity types you care about at inference time, so there is no fixed label schema to work around.

Fine-tuned from urchade/gliner_small-v2.1 on the full Nemotron-PII dataset (~100k).

Results

Held-out test set (100 examples), 24 fine-grained PII labels:

Metric	F1
Fine-grained	92.8%
Coarse-grained	94.4%

NVIDIA's Nemotron-PII reference: fine 96.2%, coarse 96.7%.

Usage

from gliner import GLiNER

model = GLiNER.from_pretrained("daslabhq/pii-proxy")
labels = ["first_name", "last_name", "email", "phone_number", "ssn", "street_address"]
entities = model.predict_entities("Patient Marcus Weber, marcus.weber@gmail.com", labels)
for e in entities:
    print(e["text"], "->", e["label"])

Training

• Base model: urchade/gliner_small-v2.1
• Epochs: 5, batch size 16
• Train examples: 98215
• Focal loss (alpha 0.75, gamma 2)