privaset / README.md
akaruineko's picture
Update README.md
d07125d verified
|
Raw
History Blame Contribute Delete
3.79 kB
metadata
license: mit
language:
  - zh
  - ar
  - hi
  - en
  - es
  - bn
  - pt
  - ru
  - ja
  - de
  - ko
  - fr
  - jv
  - te
  - mr
  - vi
  - ta
  - it
  - tr
  - ur
  - pa
  - uk
  - gu
  - th
  - pl
tags:
  - pii
  - private
  - pii-detection
  - privacy
pretty_name: PRIVAset
size_categories:
  - 1M<n<10M

PRIVAset: A Multilingual Synthetic PII Detection Dataset

PRIVAset is a large-scale, privacy-safe, synthetic dataset for training and evaluating Personally Identifiable Information (PII) filtering systems. It supports 25 languages, includes 16 PII types, and is designed for tasks such as classification, named-entity recognition (NER), and redaction.

All data is generated using Faker (locale-aware) and custom synthetic logic. No real PII is used, making it safe for public release, model fine-tuning, and benchmarking.


Dataset Highlights

  • 16 PII types – including PERSON_NAME, EMAIL, PHONE, SSN, CREDIT_CARD, IPV4/6, STREET_ADDRESS, DATE_OF_BIRTH, PASSPORT, DRIVERS_LICENSE, IBAN, TAX_ID, API_KEY, MEDICAL_RECORD, and BANK_ACCOUNT.
  • 25 languages – with native-script templates and locale-aware Faker support. Languages range from high-resource (English, Chinese, Arabic) to lower-resource (Javanese, Gujarati, Punjabi).
  • 10 real-world contexts – including chat, email, support tickets, medical notes, legal documents, code snippets, social posts, resumes, ecommerce, and finance.
  • Balanced positives and negatives – with hard negatives that look like PII but are not.
  • Token‑level annotations – each record includes character-level spans (start, end, type, value), document‑level label (has_pii), and risk score (high/medium/low).
  • Multi‑PII examples – ~30% of records contain 2–4 PII entities.
  • Robustness augmentations – including email obfuscation ([at]/(dot)), phone format variance, and case flipping.
  • Flexible exports – available in JSONL, CSV, Hugging Face datasets, and instruction‑tuning formats (prompt/completion).

Supported Languages

Language Code Faker Locale Native Script
Chinese zh zh_CN 中文
Arabic ar ar_SA العربية
Hindi hi hi_IN हिन्दी
English en en_US English
Spanish es es_ES Español
Bengali bn bn_BD বাংলা
Portuguese pt pt_PT Português
Russian ru ru_RU Русский
Japanese ja ja_JP 日本語
German de de_DE Deutsch
Korean ko ko_KR 한국어
French fr fr_FR Français
Javanese jv id_ID Basa Jawa
Telugu te en_IN తెలుగు
Marathi mr mr_IN मराठी
Vietnamese vi vi_VN Tiếng Việt
Tamil ta ta_IN தமிழ்
Italian it it_IT Italiano
Turkish tr tr_TR Türkçe
Urdu ur en_PK اردو
Punjabi pa en_IN ਪੰਜਾਬੀ
Ukrainian uk uk_UA Українська
Gujarati gu gu_IN ગુજરાતી
Thai th th_TH ภาษาไทย
Polish pl pl_PL Polski

Use Cases

  • Multilingual PII classification – use has_pii + language for per‑language calibration.
  • NER / Token classification – use pii spans to generate BIO tags; spans are character‑accurate for CJK and Arabic scripts.
  • Redaction systems – replace spans with [REDACTED_{TYPE}].
  • Risk‑aware filtering – block high‑risk entities (SSN, credit card, IBAN, API key) and warn on medium risk.

License

This dataset is released under the MIT License.