---
license: apache-2.0
language:
  - en
  - nl
  - de
  - fr
  - es
  - it
pipeline_tag: token-classification
base_model: microsoft/Multilingual-MiniLM-L12-H384
datasets:
  - ai4privacy/pii-masking-openpii-1.5m
  - Wismut/nym-pii-multilingual-data
tags:
  - pii
  - privacy
  - ner
  - onnx
---

# Temperate PII

This is the model of [Temperate PII](https://github.com/temperate-works/pii), a service that finds personal data in European text. The model finds person names and street addresses. Rules per country find the structured values, such as IBANs, phone numbers and national ID numbers, and check their checksums.

The model is a 6-layer student, distilled from a larger teacher and built on Multilingual MiniLM. It has 107 million parameters and runs as an int8 ONNX model on a CPU. It was trained on English, Dutch, German, French, Spanish and Italian.

## Results

On the benchmark in the GitHub repository, 2,041 texts in the six languages:

| Type | Values | Recall | Precision |
| --- | ---: | ---: | ---: |
| Person names | 205 | 98.0% | 96.2% |
| Street addresses | 275 | 100.0% | 94.2% |

On 26 business texts that we wrote by hand, such as emails, invoices and contracts, it found 76 of the 83 names and addresses. We built the benchmark ourselves, so read these numbers with care. The repository explains how we measured, and has the results for every type and a comparison with Azure AI Language.

## Use

Run the service from the repository. `fetch-model.sh` downloads these files and checks each one against its digest:

```bash
git clone https://github.com/temperate-works/pii
cd pii
./fetch-model.sh
```

| File | What it holds |
| --- | --- |
| `model.onnx` | The model, int8 with static quantization |
| `tokenizer.json` | The tokenizer of `FacebookAI/xlm-roberta-base` |
| `labels.json` | The BIO labels. The service uses only `PERSON` and `STREET_ADDRESS`; rules find the other types |
| `calibration.json` | The score threshold per type |

The service reads texts in windows of 256 tokens that overlap by 32.

A surname on its own is left alone on purpose, when no title, greeting, label or earlier full name points to a person. Masking every word that could be a surname would also mask many company and place names.

## Licence

Apache-2.0. The model is built on `microsoft/Multilingual-MiniLM-L12-H384` (MIT) and the tokenizer of `FacebookAI/xlm-roberta-base` (MIT). It was trained on `ai4privacy/pii-masking-openpii-1.5m` by ai4privacy (CC BY 4.0) and `Wismut/nym-pii-multilingual-data` (MIT).

Made by [Temperate](https://temperate.works).