GolemGuard: Hebrew Privacy Information Detection Corpus

GolemGuard is a comprehensive Hebrew language dataset specifically designed for training and evaluating models for Personal Identifiable Information (PII) detection and masking. The dataset contains ~600MB of synthetic text data representing various document types and communication formats commonly found in Israeli professional and administrative contexts.

Source Data

Initial Data Collection and Normalization

The dataset combines synthetic data generated from multiple authoritative sources:

1. Names:

   • Israeli government open data portal (data.gov.il) for first names
   • Hebrew Wikipedia category for family names
   • Faker library
   • Additional curated sources

2. Addresses:

   • Israeli government datasets for cities
   • Israeli government street names datasets

3. Synthetic Identifiers:

   • ID numbers: Generated following Israeli ID number format and checksum rules
   • Bank accounts: Generated following Israeli bank account number formats
   • Postal codes: Based on valid Israeli postal code ranges
   • Credit cards: Generated following valid card number algorithms
   • Email addresses: Constructed using collected names and top Israeli email providers

Entity Types:

PII Entity Type	Description	Count
FIRST_NAME	First names	131,192
LAST_NAME	Last names	115,020
ID_NUM	Israeli ID numbers	77,902
PHONE_NUM	Phone numbers (both mobile and landlines)	142,795
DATE	Dates	77,042
STREET	Street addresses	105,381
CITY	City names	105,874
EMAIL	Email addresses	90,080
POSTAL_CODE	Israeli postal codes (Mikud)	93,748
BANK_ACCOUNT_NUM	Bank account numbers	31,046
CC_NUM	Credit card numbers	1,751
CC_PROVIDER	Credit card providers	1,679

Template Diversity

The dataset includes 2,607 unique document templates. Here are the top document types by instance count:

Template Type	Instance Count
meeting_summary	21,983
appointment_confirmation	4,420
job_application_confirmation	3,611
job_application	3,306
medical_appointment_confirmation	2,359
loan_application_received	1,988
job_application_received	1,832
bank_account_update	1,609
appointment_reminder	1,567
medical_appointment_reminder	1,335

Dataset Size

• Total Size: ~600MB
• Number of Examples: 115,453
• Format: JSONL

Data Splits

Split	Number of Instances
Training	97,453
Test	18,000
Total	115,453

Languages

• Hebrew (he)
• Locale: Israel (IL)

Supported Tasks

• Token Classification
• Named Entity Recognition (NER)
• PII Detection
• Text Masking
• Privacy-Preserving Text Processing

Dataset Structure

Data Instances

Each instance in the dataset contains:

{
  "id": "String",                    // Unique identifier
  "source_text": "String",          // Original text with PII
  "masked_text": "String",          // Text with PII entities masked
  "locale": "String",              // Always "IL"
  "language": "String",            // Always "he"
  "split": "String",              // "train" or "test"
  "privacy_mask": [               // List of PII entities with positions
    {
      "label": "String",         // Entity type
      "start": "Integer",       // Start position
      "end": "Integer",        // End position
      "value": "String",      // Original value
      "label_index": "Integer" // Index for multiple entities of same type
    }
  ],
  "span_labels": "List",        // Entity spans in [start, end, label] format
  "tokens": "List",            // Tokenized text
  "token_classes": "List",    // Token-level BIO tags
  "input_ids": "List",       // Model input token IDs
  "attention_mask": "List",  // Attention mask for padding
  "offset_mapping": "List", // Character offsets for tokens
  "template_type": "String" // Document type/template
}

Sample

{"id": "entry_000256", "source_text": "משתתף: דולב שם טוב\nתאריך לידה: 28.08.97\nכתובת: סחרוב דוד 158, קריית ים, 2676389\nאימייל: sagitbendor760@live.com\nטלפון: +972-58-2892208\n\nסיכום הפגישה: בפגישה זו נדונה חשיבות שיפור התקשורת הפנימית בין הצוותים. הוחלט לבצע סדנאות להכשרה נוספת בשבוע הבא.", "masked_text": "משתתף: [FIRST_NAME_1] [LAST_NAME_1]\nתאריך לידה: [DATE_1]\nכתובת: [STREET_1], [CITY_1], [POSTAL_CODE_1]\nאימייל: [EMAIL_1]\nטלפון: [PHONE_NUM_1]\n\nסיכום הפגישה: בפגישה זו נדונה חשיבות שיפור התקשורת הפנימית בין הצוותים. הוחלט לבצע סדנאות להכשרה נוספת בשבוע הבא.", "locale": "IL", "language": "he", "split": "train", "privacy_mask": [{"label": "FIRST_NAME", "start": 7, "end": 11, "value": "דולב", "label_index": 1}, {"label": "LAST_NAME", "start": 12, "end": 18, "value": "שם טוב", "label_index": 1}, {"label": "DATE", "start": 31, "end": 39, "value": "28.08.97", "label_index": 1}, {"label": "STREET", "start": 47, "end": 60, "value": "סחרוב דוד 158", "label_index": 1}, {"label": "CITY", "start": 62, "end": 70, "value": "קריית ים", "label_index": 1}, {"label": "POSTAL_CODE", "start": 72, "end": 79, "value": "2676389", "label_index": 1}, {"label": "EMAIL", "start": 88, "end": 111, "value": "sagitbendor760@live.com", "label_index": 1}, {"label": "PHONE_NUM", "start": 119, "end": 134, "value": "+972-58-2892208", "label_index": 1}], "span_labels": [[7, 11, "FIRST_NAME"], [12, 18, "LAST_NAME"], [31, 39, "DATE"], [47, 60, "STREET"], [62, 70, "CITY"], [72, 79, "POSTAL_CODE"], [88, 111, "EMAIL"], [119, 134, "PHONE_NUM"]], "tokens": ["<s>", "▁מ", "שתתף", ":", "▁דו", "לב", "▁שם", "▁טוב", "▁", "תאריך", "▁", "לידה", ":", "▁28.", "08.", "97", "▁כתובת", ":", "▁", "סחר", "וב", "▁דוד", "▁158", ",", "▁קרי", "ית", "▁", "ים", ",", "▁26", "76", "389", "▁אימייל", ":", "▁sa", "git", "ben", "dor", "760", "@", "live", ".", "com", "▁טלפון", ":", "▁+", "97", "2-", "58", "-28", "92", "208", "▁", "סיכום", "▁הפ", "גישה", ":", "▁בפ", "גישה", "▁זו", "▁", "נד", "ונה", "▁", "חשיבות", "▁", "שיפור", "▁התקשורת", "▁הפנימי", "ת", "▁בין", "▁הצוות", "ים", ".", "▁ה", "וח", "לט", "▁לבצע", "▁", "סדנאות", "▁ל", "הכשרה", "▁נוספת", "▁בשבוע", "▁הבא", ".", "</s>"], "token_classes": ["O", "O", "O", "O", "B-FIRST_NAME", "I-FIRST_NAME", "B-LAST_NAME", "I-LAST_NAME", "O", "O", "O", "O", "O", "B-DATE", "I-DATE", "I-DATE", "O", "O", "B-STREET", "I-STREET", "I-STREET", "I-STREET", "I-STREET", "O", "B-CITY", "I-CITY", "I-CITY", "I-CITY", "O", "B-POSTAL_CODE", "I-POSTAL_CODE", "I-POSTAL_CODE", "O", "O", "B-EMAIL", "I-EMAIL", "I-EMAIL", "I-EMAIL", "I-EMAIL", "I-EMAIL", "I-EMAIL", "I-EMAIL", "I-EMAIL", "O", "O", "B-PHONE_NUM", "I-PHONE_NUM", "I-PHONE_NUM", "I-PHONE_NUM", "I-PHONE_NUM", "I-PHONE_NUM", "I-PHONE_NUM", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O"], "input_ids": [0, 874, 177834, 12, 18133, 22849, 16804, 19767, 6, 174081, 6, 119583, 12, 12511, 17331, 14773, 86185, 12, 6, 214797, 18340, 67512, 78373, 4, 46050, 2754, 6, 448, 4, 1381, 11835, 119861, 194921, 12, 57, 15769, 776, 1846, 110216, 981, 24056, 5, 277, 167153, 12, 997, 14773, 18504, 10057, 48590, 12231, 154782, 6, 186708, 18338, 58366, 12, 65412, 58366, 8248, 6, 16747, 15703, 6, 148055, 6, 154997, 185983, 214547, 609, 7501, 192819, 448, 5, 364, 15662, 21295, 107543, 6, 199930, 657, 226909, 122485, 132299, 54641, 5, 2], "attention_mask": [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], "offset_mapping": [[0, 0], [0, 1], [1, 5], [5, 6], [7, 9], [9, 11], [12, 14], [15, 18], [19, 20], [19, 24], [25, 26], [25, 29], [29, 30], [31, 34], [34, 37], [37, 39], [40, 45], [45, 46], [47, 48], [47, 50], [50, 52], [53, 56], [57, 60], [60, 61], [62, 65], [65, 67], [68, 69], [68, 70], [70, 71], [72, 74], [74, 76], [76, 79], [80, 86], [86, 87], [88, 90], [90, 93], [93, 96], [96, 99], [99, 102], [102, 103], [103, 107], [107, 108], [108, 111], [112, 117], [117, 118], [119, 120], [120, 122], [122, 124], [124, 126], [126, 129], [129, 131], [131, 134], [136, 137], [136, 141], [142, 144], [144, 148], [148, 149], [150, 152], [152, 156], [157, 159], [160, 161], [160, 162], [162, 165], [166, 167], [166, 172], [173, 174], [173, 178], [179, 186], [187, 193], [193, 194], [195, 198], [199, 204], [204, 206], [206, 207], [208, 209], [209, 211], [211, 213], [214, 218], [219, 220], [219, 225], [226, 227], [227, 232], [233, 238], [239, 244], [245, 248], [248, 249], [0, 0]], "template_type": "meeting_summary"}

Considerations for Using the Data

Social Impact of Dataset

The dataset aims to improve privacy protection in Hebrew text processing by enabling better PII detection and masking. This has important applications in:

Application Area	Description
Regulatory Compliance	Support for GDPR and PPLA requirements
Document Processing	Privacy-preserving text analysis and storage
Information Security	Automated PII detection and protection
Data Loss Prevention	Real-time PII identification and masking

Discussion of Biases

1. Geographic Bias

   • Dataset focuses on Israeli context and formats

2. Name Distribution

   • While effort was made to include diverse names, distribution may not perfectly match population demographics

Additional Information

Dataset Curators

Dataset was generated by Liran Baba.

Model Training

Dataset was generated for training the GolemPII-xlm-roberta-v1 model by Liran Baba (CordwainerSmith), model, demonstrating its effectiveness for PII detection and masking tasks in Hebrew text.

Recommended Uses

Use Case	Description
Privacy Protection	Identifying and masking PII in Hebrew documents
Compliance Checking	Automated PII detection for regulatory compliance
Data Sanitization	Cleaning sensitive information from text data
Information Security	Supporting data loss prevention systems

Quick Start

from datasets import load_dataset

# Load dataset
dataset = load_dataset("GolemGuard")

# Example usage
sample = dataset['train'][0]
print(f"Original text: {sample['source_text']}")
print(f"Masked text: {sample['masked_text']}")

Versioning

This is the initial release of the dataset. Future versions may include:

• Additional template types
• Expanded entity coverage
• Enhanced demographic representation
• Additional language variants

License

The GolemGuard dataset 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 dataset in your research, project, or application, please include the following citation:

For informal usage (e.g., blog posts, documentation):

GolemGuard Dataset by Liran Baba (https://huggingface.co/datasets/CordwainerSmith/GolemGuard)

For academic or professional publications:

Baba, L. (2024). GolemGuard: A Professional Hebrew PII Detection Dataset. 
Retrieved from https://huggingface.co/datasets/CordwainerSmith/GolemGuard

Related model: GolemPII-xlm-roberta-v1 (https://huggingface.co/CordwainerSmith/GolemPII-xlm-roberta-v1)

Usage Examples

When referencing in your code:

"""
This code uses the GolemGuard dataset by Liran Baba
(https://huggingface.co/datasets/CordwainerSmith/GolemGuard)
"""

from datasets import load_dataset
dataset = load_dataset("GolemGuard")

When referencing in your model card:

dataset_info:
  - name: GolemGuard
    author: Liran Baba
    url: https://huggingface.co/datasets/CordwainerSmith/GolemGuard
    year: 2024