SemplificaAI/gliner2-privacy-filter-PII-multi

GLiNER2 Privacy-Filter PII Multi (ONNX Fragmented & IOBinding)

This repository contains the ONNX-exported weights of fastino/gliner2-privacy-filter-PII-multi, the multilingual PII detection model built on GLiNER2 by Fastino AI.

The model is exported in a fragmented format (encoder, token_gather, span_rep, schema_gather, count_pred_argmax, count_lstm_fixed, scorer, classifier) for direct compatibility with gliner2-rs, the official Zero-Python Native Rust inference engine for GLiNER2.

It supports detection of 42 PII entity types across 7 languages (EN, FR, ES, DE, IT, PT, NL).

---

🆕 V2 Zero-Copy IOBinding Models

Like the gliner2-multi-v1-onnx base release, this repo ships the V2 fused IOBinding variant. Gather, ArgMax, MatMul operations are fused directly into the ONNX graphs so that tensors never leave the GPU/NPU VRAM, bypassing the PCIe bus and cutting inference latency by ~30 % on discrete GPUs.

📂 Available Variants

Variant	Use case	Notes
fp16_v2 (recommended)	NVIDIA CUDA · AMD ROCm · Apple CoreML · Qualcomm QNN	Zero-Copy VRAM (IOBinding), full FP16 IO, fused ops
fp32_v2	CPU (AVX2 / XNNPACK / ARM NEON)	High precision V2 fusions for CPU
fp16 (standard)	Legacy compatible, all EPs	FP32 IO (CoreML-compatible), slower on CUDA due to PCIe round-trips
fp32 (standard)	Universal fallback	Legacy Float32

Each variant ships 8 fragments:

encoder_{precision}.onnx          ~530–1060 MB
token_gather_{precision}.onnx     ~ <1 MB
span_rep_{precision}.onnx         ~32–63 MB
schema_gather_{precision}.onnx    ~ <1 MB
count_pred_argmax_{precision}.onnx ~2–5 MB
count_lstm_fixed_{precision}.onnx ~20–41 MB
scorer_{precision}.onnx           ~ <1 MB
classifier_{precision}.onnx       ~2–5 MB

Total: ~590 MB (FP16) or ~1.17 GB (FP32) per variant.

---

🎯 Supported PII Labels (42 types)

Person / Names (6 labels)

person, full_name, first_name, middle_name, last_name, date_of_birth

Contact / Address (8 labels)

email, phone_number, address, street_address, city, state_or_region, postal_code, country

Government / Tax IDs (7 labels)

government_id, national_id_number, passport_number, drivers_license_number, license_number, tax_id, tax_number

Banking / Payment (8 labels)

bank_account, account_number, routing_number, iban, payment_card, card_number, card_expiry, card_cvv

Digital Identity (4 labels)

username, ip_address, account_id, sensitive_account_id

Secrets / Credentials (5 labels)

password, secret, api_key, access_token, recovery_code

Sensitive Dates (4 labels)

sensitive_date, document_date, expiration_date, transaction_date

---

🚀 Usage in Rust (gliner2-rs)

use gliner2_inference::{Gliner2Engine, ModelType, SchemaTask};

// Auto-downloads the V2 FP16 fragments from this HuggingFace repo
// and switches to the high-performance IOBinding engine.
let engine = Gliner2Engine::from_pretrained(
    "SemplificaAI/gliner2-privacy-filter-PII-multi",
    Some("fp16_v2"),
    ModelType::HuggingFace,
)?;

let text = "Please contact Maria Jensen at maria.jensen@example.dk or +45 20 12 34 56.";
let tasks = vec![
    SchemaTask::Entities(vec![
        "person".into(), "email".into(), "phone_number".into(),
    ])
];

let (entities, _, _) = engine.extract(text, &tasks)?;

Requires gliner2-rs >= 0.4.1 for automatic V2 detection / IOBinding routing.

🐍 Usage in Python (onnxruntime)

Run the 8-fragment pipeline manually (no Python gliner2 dependency needed):

import onnxruntime as ort

# Per fragment (example for the encoder, CUDA backend)
encoder = ort.InferenceSession(
    "encoder_fp16_iobinding.onnx",
    providers=["CUDAExecutionProvider"],
)
# ...load the other 7 fragments analogously...

# Chain them via IOBinding (see validate_onnx_v2.py for a full reference impl)

For a simpler entry point you can keep using the original PyTorch model via the gliner2 Python package on fastino/gliner2-privacy-filter-PII-multi; this ONNX repo is optimised for production deployment without Python.

---

🛠 Pipeline Wiring (IOBinding chain)

encoder_fp16_iobinding.onnx
    │
    ├─ token_gather_fp16_iobinding.onnx
    │       └─ span_rep_fp16_iobinding.onnx
    │
    └─ schema_gather_fp16_iobinding.onnx
            ├─ count_pred_argmax_fp16_iobinding.onnx  →  pred_count (int64)
            └─ count_lstm_fixed_fp16_iobinding.onnx
                    └─ scorer_fp16_iobinding.onnx     →  entity_scores

classifier_fp16_iobinding.onnx (only for classification tasks)

---

⚙️ Technical Notes

• opset 17 (ONNX 1.14+) for maximum execution-provider compatibility.
• count_lstm_fixed exports the GRU unrolled to 20 fixed steps at tracing time → compatible with execution providers that don't support dynamic loops (Apple CoreML, Qualcomm QNN).
• scorer uses fused Reshape + MatMul + Transpose instead of Einsum for compatibility with QNN/CoreML FP16.
• INT8 not supported: the DeBERTa-v3 disentangled-attention activations contain extreme outliers that saturate 8-bit ranges (the same limitation called out by the GLiNER2 maintainers). FP16 remains the optimal compression target.
• Encoder size: ~1.06 GB FP32 → ~530 MB FP16. Larger than the multi-v1 base because of the wider classification head (42 PII labels) and per-language fine-tuning.

🪪 License

Apache 2.0 — same as the upstream model.

🙏 Acknowledgements

• Upstream model: fastino/gliner2-privacy-filter-PII-multi by Fastino AI.
• GLiNER2 paper: Zaratiana et al., GLiNER2: Schema-Driven Multi-Task Learning for Structured Information Extraction, EMNLP 2025.
• ONNX fragmentation + IOBinding strategy: Semplifica s.r.l., as used in gliner2-multi-v1-onnx.

📚 Citation

@misc{fastino2026gliner2pii,
  title   = {GLiNER2-PII: Multilingual PII Extraction via Synthetic Fine-Tuning},
  author  = {{Fastino AI Team}},
  year    = {2026},
  url     = {https://huggingface.co/fastino/gliner2-privacy-filter-PII-multi}
}

@inproceedings{zaratiana-etal-2025-gliner2,
  title     = {GLiNER2: Schema-Driven Multi-Task Learning for Structured Information Extraction},
  author    = {Zaratiana, Urchade and Pasternak, Gil and Boyd, Oliver and Hurn-Maloney, George and Lewis, Ash},
  booktitle = {Proceedings of EMNLP 2025: System Demonstrations},
  year      = {2025}
}