File size: 6,538 Bytes
708b3d0
 
 
5f2324a
708b3d0
 
 
 
 
 
5f2324a
 
 
708b3d0
5f2324a
708b3d0
5f2324a
708b3d0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5f2324a
708b3d0
5f2324a
708b3d0
 
 
5f2324a
708b3d0
 
5f2324a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
708b3d0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5f2324a
 
 
 
708b3d0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5f2324a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
---
language: he
license: mit
library_name: transformers
tags:
- hebrew
- ner
- pii-detection
- token-classification
- xlm-roberta
- privacy
- data-anonymization
- golemguard
datasets:
- CordwainerSmith/GolemGuard 
model-index:
- name: GolemPII-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-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 to achieve high accuracy in identifying and classifying various types of PII.

## Model Details
- Based on xlm-roberta-base
- Fine-tuned on the GolemGuard: Hebrew Privacy Information Detection Corpus
- Optimized for token classification tasks in Hebrew text

## Intended Uses & Limitations

This model is intended for:

* **Privacy Protection:**  Detecting and masking PII in Hebrew text to protect individual privacy.
* **Data Anonymization:** Automating the process of de-identifying Hebrew documents in legal, medical, and other sensitive contexts.
* **Research:**  Supporting research in Hebrew natural language processing and PII detection.

## Training Parameters

* **Batch Size:** 32
* **Learning Rate:** 2e-5 with linear warmup and decay.
* **Optimizer:** AdamW
* **Hardware:** Trained on a single NVIDIA A100GPU. 

## Dataset Details

* **Dataset Name:** GolemGuard: Hebrew Privacy Information Detection Corpus 
* **Dataset Link:** [https://huggingface.co/datasets/CordwainerSmith/GolemGuard](https://huggingface.co/datasets/CordwainerSmith/GolemGuard)

## 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 |

## Model Architecture

The model is based on the `FacebookAI/xlm-roberta-base` architecture, a transformer-based language model pre-trained on a massive multilingual dataset.  No architectural modifications were made to the base model during fine-tuning.

## 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 = "砖诇讜诐, 砖诪讬 讚讜讚 讻讛谉 讜讗谞讬 讙专 讘专讞讜讘 讛专爪诇 42 讘转诇 讗讘讬讘. 讛讟诇驻讜谉 砖诇讬 讛讜讗 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}")
```


## License

The GolemPII-v1 model is released under MIT License with the following additional terms:

```
MIT License

Copyright (c) 2024 Liran Baba

Permission is hereby granted, free of charge, to any person obtaining a copy
of this dataset and associated documentation files (the "Dataset"), to deal
in the Dataset without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Dataset, and to permit persons to whom the Dataset is
furnished to do so, subject to the following conditions:

1. The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Dataset.

2. Any academic or professional work that uses this Dataset must include an 
appropriate citation as specified below.

THE DATASET IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE DATASET OR THE USE OR OTHER DEALINGS IN THE
DATASET.
```

### How to Cite

If you use this model in your research, project, or application, please include the following citation:

For informal usage (e.g., blog posts, documentation):
```
GolemPII-v1 model by Liran Baba (https://huggingface.co/CordwainerSmith/GolemPII-v1)
```