ONNX export of nvidia/gliner-PII, packaged for local, dependency-light inference. This is the on-device PII detector that ships inside Sovraine Guard, where it powers the privacy router: detection, classification, and redaction of personal data in AI-agent tool traffic — across 11 languages, entirely on-device.

Model details


Architecture	GLiNER (mdeberta-v3-base backbone) fine-tuned for PII detection
Task	Zero-shot token classification — detect arbitrary PII entity types by label name
Format	ONNX variants: FP32, FP16, INT8 (pick per latency/quality budget)
Languages	en, fr, de, es, it, pt, nl, pl, ja, ko, zh
License	NVIDIA Open Model License
Base model	nvidia/gliner-PII

Variants

File	Precision	Use case
`onnx/model.onnx`	FP32	Reference quality
`onnx/model_fp16.onnx`	FP16	Default desktop inference
`onnx/model_int8.onnx`	INT8	Constrained environments

Usage

GLiNER is zero-shot: you pass the entity labels you want detected.

python

from gliner import GLiNER

model = GLiNER.from_pretrained(
    "Sovraine/gliner-pii-onnx",
    load_onnx_model=True,
    onnx_model_file="onnx/model_fp16.onnx",
)

labels = ["person", "email", "phone", "credit card",
          "national ID", "address"]

text = "Contact Marie Dubois at marie.dubois@example.fr or +33 6 12 34 56 78."
for ent in model.predict_entities(text, labels):
    print(f"{ent['label']:>20}: {ent['text']}")
#         person name: Marie Dubois
#               email: marie.dubois@example.fr
#        phone number: +33 6 12 34 56 78

Evaluation

For benchmark results, see the upstream nvidia/gliner-PII model card. The ONNX export is a format conversion; no fine-tuning was performed by Sovraine. Quantized variants (INT8, Q4) trade some recall for footprint — validate on your own data before choosing a variant for regulated workloads.

Limitations

Zero-shot quality varies by label phrasing — descriptive labels ("credit card number") outperform terse ones ("cc")
Quantization impact — the INT8 variant may miss borderline entities that FP16/FP32 catch
Language coverage is uneven — the 11 listed languages are supported; quality is strongest in high-resource ones
Detection is probabilistic: for regulated data flows, combine with pattern rules and fail-closed policy (as Sovraine Guard does) rather than relying on NER alone

Attribution

Original model: NVIDIA — gliner-PII (NVIDIA Open Model License)
GLiNER architecture: urchade/GLiNER
ONNX packaging and distribution: Sovraine