---
language: he
license: mit
tags:
- hebrew
- ner
- pii-detection
- token-classification
- xlm-roberta
datasets:
- custom
model-index:
- name: GolemPII-xlm-roberta-v1
  results:
  - task:
      name: Token Classification
      type: token-classification
    metrics:
      - name: F1
        type: f1
        value: 0.9982
      - name: Precision
        type: precision
        value: 0.9982
      - name: Recall
        type: recall
        value: 0.9982
---

# GolemPII-xlm-roberta-v1 - Hebrew PII Detection Model

This model is trained to detect personally identifiable information (PII) in Hebrew text. While based on the multilingual XLM-RoBERTa model, it has been specifically fine-tuned on Hebrew data.

## Model Details
- Based on xlm-roberta-base
- Fine-tuned on a custom Hebrew PII dataset
- Optimized for token classification tasks in Hebrew text

## Performance Metrics

### Final Evaluation Results
```
eval_loss: 0.000729
eval_precision: 0.9982
eval_recall: 0.9982
eval_f1: 0.9982
eval_accuracy: 0.999795
```

### Detailed Performance by Label

| Label            | Precision | Recall  | F1-Score | Support |
|------------------|-----------|---------|----------|---------|
| BANK_ACCOUNT_NUM | 1.0000    | 1.0000  | 1.0000   | 4847    |
| CC_NUM          | 1.0000    | 1.0000  | 1.0000   | 234     |
| CC_PROVIDER     | 1.0000    | 1.0000  | 1.0000   | 242     |
| CITY            | 0.9997    | 0.9995  | 0.9996   | 12237   |
| DATE            | 0.9997    | 0.9998  | 0.9997   | 11943   |
| EMAIL           | 0.9998    | 1.0000  | 0.9999   | 13235   |
| FIRST_NAME      | 0.9937    | 0.9938  | 0.9937   | 17888   |
| ID_NUM          | 0.9999    | 1.0000  | 1.0000   | 10577   |
| LAST_NAME       | 0.9928    | 0.9921  | 0.9925   | 15655   |
| PHONE_NUM       | 1.0000    | 0.9998  | 0.9999   | 20838   |
| POSTAL_CODE     | 0.9998    | 0.9999  | 0.9999   | 13321   |
| STREET          | 0.9999    | 0.9999  | 0.9999   | 14032   |
| micro avg       | 0.9982    | 0.9982  | 0.9982   | 135049  |
| macro avg       | 0.9988    | 0.9987  | 0.9988   | 135049  |
| weighted avg    | 0.9982    | 0.9982  | 0.9982   | 135049  |

### Training Progress

| Epoch | Training Loss | Validation Loss | Precision | Recall  | F1       | Accuracy |
|-------|--------------|-----------------|-----------|---------|----------|----------|
| 1     | 0.005800    | 0.002487       | 0.993109  | 0.993678| 0.993393 | 0.999328 |
| 2     | 0.001700    | 0.001385       | 0.995469  | 0.995947| 0.995708 | 0.999575 |
| 3     | 0.001200    | 0.000946       | 0.997159  | 0.997487| 0.997323 | 0.999739 |
| 4     | 0.000900    | 0.000896       | 0.997626  | 0.997868| 0.997747 | 0.999750 |
| 5     | 0.000600    | 0.000729       | 0.997981  | 0.998191| 0.998086 | 0.999795 |

## Usage
```python
import torch
from transformers import AutoTokenizer, AutoModelForTokenClassification

tokenizer = AutoTokenizer.from_pretrained("{repo_id}")
model = AutoModelForTokenClassification.from_pretrained("{repo_id}")

# Example text (Hebrew)
text = "שלום, שמי דוד כהן ואני גר ברחוב הרצל 42 בתל אביב. הטלפון שלי הוא 050-1234567"

# Tokenize and get predictions
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
with torch.no_grad():
    outputs = model(**inputs)
    predictions = torch.argmax(outputs.logits, dim=2)

# Convert predictions to labels
tokens = tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])
labels = [model.config.id2label[t.item()] for t in predictions[0]]

# Print results (excluding special tokens and non-entity labels)
for token, label in zip(tokens, labels):
    if label != "O" and not token.startswith("##"):
        print(f"Token: {token}, Label: {label}")
```

## Training Details
- Training epochs: 5
- Training speed: ~2.33 it/s (7615/7615 54:29)
- Base model: xlm-roberta-base
- Training language: Hebrew