ph-eye-pii-en-xsmall

A GLiNER person-name detector for English PII redaction, fine-tuned by Philterd from microsoft/deberta-v3-xsmall on nvidia/Nemotron-PII. It detects a single label, name, which is what PhEye consumes for redaction.

This is the ultra-light variant — the smallest size, for the most constrained on-device / edge / in-browser use (about half the parameters and latency of the small model). It ships a quantized int8 ONNX graph (about 90 MB) and runs in-process with no server. Sibling sizes share the same recipe and data: philterd/ph-eye-pii-en-small, philterd/ph-eye-pii-en-medium (the recommended default), and philterd/ph-eye-pii-en-large (highest-capacity, server-side).

Use it in production

This model only detects names. Turning detections into redacted output (policy-driven replacement, masking, encryption, format-preserving encryption, PDF handling, and hosting) is what Philter and PhEye add around it. PhEye serves this model behind a simple API; Philter wraps it in a full self-hosted PII/PHI redaction pipeline you run inside your own boundary.

Commercial support, custom model training, and consulting are available from Philterd, including production support and custom redaction engagements.

Intended use

Detecting personal names in English text so they can be redacted, where footprint and latency matter most. The model is recall-leaning by design: in redaction, missing a name (a leak) is worse than flagging an extra span (over-redaction). We recommend a default threshold of 0.50 — the operating point the evaluation below reports. Lower it to push recall higher at the cost of more over-redaction.

Because it is the smallest size, its real-text recall is lower than the larger variants' — if a missed name is costly and footprint allows, prefer small/medium/large. For the most constrained deployments where this is the right size, validate the operating point on your own data.

Why names only? Names are the one PII type that genuinely needs a model: they have no fixed format, overlap with ordinary words, and depend on context. The other common types — emails, phone numbers, SSNs, credit cards, IP addresses — follow regular patterns and are caught more reliably (and far more cheaply) with regexes, checksums, or dictionary lookups. Spending a neural model on those would be overkill, so this model focuses on the hard part and leaves the structured types to the pattern-based detectors in PhEye.

Usage

from gliner import GLiNER

# PyTorch
model = GLiNER.from_pretrained("philterd/ph-eye-pii-en-xsmall")
text = "Please forward the invoice to Toni Levine and copy Maria Gonzalez."
for ent in model.predict_entities(text, ["name"], threshold=0.5):
    print(ent["text"], "->", ent["label"], round(ent["score"], 2))

# or the bundled int8 ONNX graph
onnx = GLiNER.from_pretrained("philterd/ph-eye-pii-en-xsmall",
                              load_onnx_model=True, onnx_model_file="onnx/model.onnx")

The directory also ships spm.model so SentencePiece-based ONNX runtimes (e.g. .NET) can load onnx/model.onnx directly.

Evaluation

Scored on the nvidia/Nemotron-PII test split (131,105 windows, long docs chunked at 1,200 chars), at the recommended threshold 0.50, name only (support 97,012).

Recall is 0.991 unrounded — in-distribution it nearly matches the larger models. These numbers are in-distribution (trained and tested on the same synthetic Nemotron distribution) and are a ceiling, not a real-document production claim.

Because the training data is synthetic, real-text recall is materially lower than these in-distribution figures — and more so for this smallest size than for the larger variants. The recommended threshold above is a synthetic operating point. For recall-first use on real-world documents, tune the threshold down — a lower cutoff accepts more over-redaction to recover recall, which is the right trade when a missed name is a leak. Always pick the operating point on a sample of your own data.

Limitations

• English only, and detects the name label only — other PII types (emails, phone numbers, IDs, addresses) are out of scope and handled elsewhere in PhEye.
• Smallest capacity, so real-document recall is lower than small/medium/large — choose this size when footprint/latency dominate; otherwise prefer a larger variant.
• Trained on synthetic data (Nemotron-PII), so real-document accuracy is lower than the in-distribution numbers above; validate on your own data.
• May miss atypical name forms (lowercase, nicknames, non-Western name structures). It is recall-leaning, so expect some over-redaction; tune the threshold for your precision/recall balance.

Training

• Base model: microsoft/deberta-v3-xsmall (MIT).
• Dataset: nvidia/Nemotron-PII (English, 100k train / 100k test, 50+ industries).
• Labels: Nemotron's first_name/last_name are merged to name (label_map), and adjacent first/last spans are merged into one full-name span, matching standard NER benchmarks and the full-name span PhEye consumes.
• Long documents are windowed at 1,200 chars during training and evaluation so names past the 512-token window are still learned and scored.
• Recipe: configs/ph-eye-pii-en-xsmall.yaml.

License and attribution

Released under CC-BY-4.0, inherited from the training data. The NVIDIA Nemotron-PII dataset is CC-BY-4.0; commercial use is permitted with attribution to NVIDIA. Any use of this model must retain that attribution. The base model microsoft/deberta-v3-xsmall is MIT.

Disclaimer

This model makes no guarantees of accuracy or completeness. Name detection is probabilistic — it will miss some names and flag some non-names, and as the smallest variant it misses more than the larger ones. Before relying on it, validate precision and recall on your own data. You remain the data controller for any personal data you process with this model and are responsible for your own compliance.