README.md

---
datasets:
- ai4privacy/pii-masking-200k
base_model:
- LH-Tech-AI/Shield-82M
pipeline_tag: token-classification
tags:
- distil
- pii
- security
- shield
- small
- cpu
- fast
- open
- open-source
- lh-tech
- bert
- roberta
library_name: transformers.js
---



# Shield-82M (ONNX)


This is an ONNX version of [LH-Tech-AI/Shield-82M](https://huggingface.co/LH-Tech-AI/Shield-82M). It was automatically converted and uploaded using [this Hugging Face Space](https://huggingface.co/spaces/onnx-community/convert-to-onnx).


## Usage with Transformers.js


See the pipeline documentation for `token-classification`: https://huggingface.co/docs/transformers.js/api/pipelines#module_pipelines.TokenClassificationPipeline


---


# 🛡️ Shield 82M
Welcome to Shield 82M, a model designed to filter PII out of texts in any language.

## Classes
This model has the following PII classes:
```plaintext
['O', 'ACCOUNTNAME', 'ACCOUNTNUMBER', 'AGE', 'AMOUNT', 'BIC', 'BITCOINADDRESS', 'BUILDINGNUMBER', 'CITY', 'COMPANYNAME', 'COUNTY', 'CREDITCARDCVV', 'CREDITCARDISSUER', 'CREDITCARDNUMBER', 'CURRENCY', 'CURRENCYCODE', 'CURRENCYNAME', 'CURRENCYSYMBOL', 'DATE', 'DOB', 'EMAIL', 'ETHEREUMADDRESS', 'EYECOLOR', 'FIRSTNAME', 'GENDER', 'HEIGHT', 'IBAN', 'IP', 'IPV4', 'IPV6', 'JOBAREA', 'JOBTITLE', 'JOBTYPE', 'LASTNAME', 'LITECOINADDRESS', 'MAC', 'MASKEDNUMBER', 'MIDDLENAME', 'NEARBYGPSCOORDINATE', 'ORDINALDIRECTION', 'PASSWORD', 'PHONEIMEI', 'PHONENUMBER', 'PIN', 'PREFIX', 'SECONDARYADDRESS', 'SEX', 'SSN', 'STATE', 'STREET', 'TIME', 'URL', 'USERAGENT', 'USERNAME', 'VEHICLEVIN', 'VEHICLEVRM', 'ZIPCODE']
```

# Base model
This model is based on distilroberta-base.

# Examples
The model has an accuracy score of ~96% (0.961206).
<br>
Here are a few examples:
### Test with name, email and phone
```plaintext
Original: My name is John Doe. Email: john@example.com. Phone: +49 123 45678.
Protected: My name is [PERSON]. Email: [EMAIL]. Phone: [PHONE].
```
### Basic test
```plaintext
Original: I live in Cambridge
Protected: I live in [ADDRESS]
```
### French test (multilingual)
```plaintext
Original: Mon e-mail est jean.dupont@example.fr et mon téléphone est +33 6 12 34 56 78.
Protected: Mon e-mail est [EMAIL] et mon téléphone est [PHONE].
```

## Quickstart
To use this model, just download `use.py` from this repo and launch it:
```bash
mkdir Shield-82M
cd Shield-82M
wget https://huggingface.co/LH-Tech-AI/Shield-82M/resolve/main/use.py
python3 use.py
```

This outputs something like:
```bash
Loading Shield-82M from LH-Tech-AI/Shield-82M...

Loading weights: 100%
 103/103 [00:00<00:00, 773.65it/s, Materializing param=roberta.encoder.layer.5.output.dense.weight]

Original: My name is John Doe. Email: john@example.com. Phone: +49 123 45678.
Protected: My name is [PERSON]. Email: [EMAIL]. Phone: [PHONE].
```

To use it with your own text, you'll have to adjust this line of code in `use.py`:
```python
sample = "My name is John Doe. Email: john@example.com. Phone: +49 123 45678."
```

## Training data
This model was trained on the first 20,000 samples of [ai4privacy/pii-masking-200k](https://huggingface.co/datasets/ai4privacy/pii-masking-200k/tree/main) for 3 epochs.

## Training details
- Epochs: 3
- Max Lenght: 512
- Base model: [distilroberta-base](https://huggingface.co/distilbert/distilroberta-base)
- Data: first 20,000 samples of [ai4privacy/pii-masking-200k](https://huggingface.co/datasets/ai4privacy/pii-masking-200k/tree/main)
- GPU: 2x Kaggle T4
- Training time: 06:38 min
- Engine: HF Transformers

The following table shows the training process:

| Epoch | Training Loss | Validation Loss | Precision | Recall | F1 | Accuracy |
| :--- | :--- | :--- | :--- | :--- | :--- | :--- |
| 1 | 1.048266 | 0.250184 | 0.904065 | 0.932844 | 0.918229 | 0.949456 |
| 2 | 0.257664 | 0.193614 | 0.939548 | 0.949651 | 0.944572 | 0.959521 |
| 3 | 0.199425 | 0.181754 | 0.939833 | 0.952215 | 0.945983 | 0.961206 |

You can find the full training code in `train.ipynb`. Runs on 2x Kaggle T4 in ~7mins.