italian-ner-pii-browser-distilbert

Italian NER for Browser-Only PII Anonymization (DistilBERT, Quantized ONNX)

A lightweight Italian Named Entity Recognition model optimized for browser-only inference, based on:

osiria/distilbert-italian-cased-ner
License: Apache-2.0
Original authors: Osiria

This repository provides a quantized ONNX version (~65 MB) suitable for running entirely in the browser via Transformers.js.

---

What this model is for

This model is designed as a privacy-friendly pre-filter layer to detect and help anonymize:

• Person names (PER)
• Organizations (ORG)
• Locations (LOC)
• Miscellaneous named entities (MISC)

Typical use case:

Run NER locally in the user's browser before sending text to an LLM, masking personal identifiers first.

All inference can run client-side.

---

Why this version

Compared to the original PyTorch model:

• Converted to ONNX
• Dynamically quantized
• Reduced payload to ~65 MB
• Optimized for ONNX Runtime Web
• Suitable for fully local browser execution

A ~65 MB download is roughly comparable to streaming 1–2 minutes of HD video (4–5 Mbps).

---

Model Details

Base model: osiria/distilbert-italian-cased-ner
Architecture: DistilBERT
Task: Token Classification (NER)
Language: Italian

This repository does not retrain the model.
It provides:

• ONNX export
• Quantized weights
• Transformers.js-compatible structure

---

Test it directly

You can test the model directly using the Hugging Face widget on this page.

Example text:

Il paziente Mario Rossi, nato il 12/04/1987, vive a Milano e lavora per l'Agenzia Spaziale Italiana.

Expected entities include:

• Mario Rossi → PER
• Milano → LOC
• Agenzia Spaziale Italiana → ORG

---

Use in the Browser (Transformers.js)

import { pipeline, env } from "@huggingface/transformers";

env.allowRemoteModels = true;
env.backends.onnx.wasm.simd = true;
env.backends.onnx.wasm.numThreads = 2;

const ner = await pipeline(
  "token-classification",
  "laibniz/italian-ner-pii-browser-distilbert",
  {
    quantized: true,
    aggregation_strategy: "simple"
  }
);

const text = "Il paziente Mario Rossi vive a Milano.";
const entities = await ner(text);
console.log(entities);

All inference runs locally in the browser.

---

Privacy & Deployment

This model is intended for:

• Browser-based anonymization layers
• On-premise deployments
• Environments where raw text must not leave the device

It should be combined with:

• Regex-based detectors (email, IBAN, VAT, tax ID, phone numbers)
• Post-processing logic
• Manual validation for high-stakes use cases

This is not a full anonymization system by itself.

---

Limitations

• Quantization may slightly reduce precision/recall
• Domain-specific texts (medical, legal, dialectal Italian) may require fine-tuning
• Token-level predictions require aggregation logic

---

Attribution

This work builds upon:

osiria/distilbert-italian-cased-ner
Apache-2.0 License
https://huggingface.co/osiria/distilbert-italian-cased-ner

All credit for original model training and dataset preparation belongs to the original authors.

This repository provides ONNX export and quantized packaging for browser use.

---

Custom Privacy-Friendly Setups

If you need:

• Domain-specific anonymization
• Clinical or legal tuning
• On-premise secure deployment
• Integration with LLM pipelines

Feel free to reach out.