The EPII PDI Masking Dataset is a large-scale, multilingual dataset of 91,400 annotated text samples containing synthetic Personal Digital Information. Each entry includes source text with embedded PII, a masked version, character-level privacy annotations, and mBERT token-level BIO tags — ready for training NER and token classification models.

Designed for enterprise applications requiring robust detection and masking of personal digital information PII, supporting compliance with GDPR, CCPA, NIS2 and other data protection regulations.

Dataset Details

Property	Value
Total Entries	91,400
Unique PII Labels	10
Languages	6 (EN, FR, DE, IT, ES, NL)
Locales	14
Split	80/20 train/validation
Format	JSON Lines (.jsonl)
Tokenization	mBERT (`bert-base-multilingual-cased`)
Annotation	Character-level spans + BIO token tags
License	All Rights Reserved

Language & Locale Distribution

Language	Entries	Locales
English	26,102	en_US, en_GB, en_CA, en_IN
French	19,543	fr_FR, fr_CA, fr_CH
Spanish	13,082	es_ES, es_MX
German	13,079	de_DE, de_CH
Italian	13,047	it_IT, it_CH
Dutch	6,547	nl_NL

PDI Label Taxonomy (10 labels)

PDI_LABELS = [
    "APIKEY",                  # API keys
    "EMAIL",                   # Email address
    "IPV4",                    # IPv4 address
    "IPV6",                    # IPv6 address
    "OTP",                     # One-time password
    "PASSWORD",                # Authentication string
    "PHONEIMEI",               # Mobile equipment identity number
    "URL",                     # Web address
    "USERAGENT",               # Browser and OS identifier
    "USERNAME",                # System identifier
]

Data Structure

Each line in the JSONL file is a JSON object with 10 fields:

{
  "source_text": "Original text with synthetic PII embedded...",
  "masked_text": "Text with PII replaced by [PDI_LABEL_N] tokens...",
  "privacy_mask": [
    {"label": "LABEL", "start": 0, "end": 10, "value": "original_value", "label_index": 1}
  ],
  "split": "train",
  "uid": 12345,
  "language": "en",
  "region": "US",
  "script": "Latn",
  "mbert_tokens": ["token1", "token2", "..."],
  "mbert_token_classes": ["O", "B-LABEL", "I-LABEL", "..."]
}

Field Reference

Field	Type	Description
`source_text`	string	Original text with synthetic PII embedded
`masked_text`	string	Text with PII replaced by `[LABEL_N]` tokens
`privacy_mask`	array	Character-level PII spans with label, position, and value
`split`	string	`"train"` or `"validation"` (80/20)
`uid`	integer	Unique entry identifier
`language`	string	ISO 639-1 language code
`region`	string	ISO 3166-1 alpha-2 region code
`script`	string	ISO 15924 script code
`mbert_tokens`	array	Multilingual BERT tokenization of source text
`mbert_token_classes`	array	BIO-tagged token classes aligned to mbert_tokens

Use Cases

PII Detection Models: Train NER / token classification models to detect personal digital information PII
Data Masking Pipelines: Build automated anonymization systems
Compliance Automation: Mask PII in logs, transcripts, emails for GDPR, CCPA, NIS2
LLM Safety: Prevent language models from exposing sensitive information
Multilingual NER: Train cross-lingual models across 6 languages and 14 locales

Citation

@dataset{ai4privacy_pdi_masking_2026,
  author = {Ai4Privacy},
  title = {EPII Personal Digital Information (PDI) Masking Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/ai4privacy/pdi-masking-100k-full}
}

License & Terms

Contact: enterprise@ai4privacy.com
Website: www.Ai4Privacy.com

This dataset contains synthetic data only — no real PII. Any system built using this data must be validated rigorously and comply with all applicable data privacy laws.

Ai4Privacy is a project affiliated with Ai Suisse SA.

p5y Data Analytics

This dataset is built on the p5y framework - think of it as i18n but for privacy. Just as i18n (internationalization) translates content into different locales, p5y translates sensitive data into privacy-safe formats through a standardized 3-step approach:

Awareness - Scan and markup private entities in unstructured text, producing a structured privacy mask with entity types, distribution, density, and risk assessment.
Protection - Control identified personal data through masking, pseudonymization, or k-anonymization, tailored to the specific use case and regulatory requirements.
Quality Assurance - Measure remaining privacy risk after anonymization, evaluating de-anonymization risks through expert annotation and automated assessment.

Learn more at p5y.org

Downloads last month: 23

Total file size:

923 MB