Loading datasets...
Release Date: 8/10/2026
Format: JSONL
Size: 5.77 MB
Share
PII Redact Urdu UOPF is a curated Urdu dataset for training and evaluating OPF privacy/PII span detection across legal, finance, social media, news/media, healthcare, education, ecommerce, telecom, government, travel, HR, and insurance domains. The dataset is built from real, non personal Urdu source text curated across these domains, into which synthetic (fabricated) PII such as names, addresses, phone numbers, emails, and account numbers was inserted and labeled with character-level spans.
Licensing
Creative Commons Attribution Non Commercial Share Alike 4.0 International (CC-BY-NC-SA-4.0)
https://spdx.org/licenses/CC-BY-NC-SA-4.0.htmlRestrictions/Special Constraints
This dataset is licensed under CC-BY-NC-SA-4.0 and is intended for non-commercial research use only, including PII detection model training and privacy-filter fine-tuning. Attribution to Proxima AI is required, and any redistributed derivative must be released under the same license.
Forbidden Usage
Use of this dataset, or of models trained on it, for commercial products or services is forbidden, consistent with the non-commercial (NC) restriction of this license.
Ethical Review
The underlying source text in this dataset is real curated Urdu content written in house by Proxima AI, containing no personal or identifying information. All PII entities present in the dataset including person names, national ID/CNIC style numbers, phone numbers, email addresses, and account numbers were synthetically fabricated and inserted for annotation purposes none correspond to real individuals. No real personal data, human subjects, or private communications were used in creating this dataset, and no third party content rights were implicated since the source text was authored in house.
Intended Use
This dataset is intended for OPF/privacy filter fine tuning, Urdu PII span detection experiments, and domain robustness checks across Urdu legal, financial, media, and operational text.
A large-scale Urdu dataset for Personally Identifiable Information (PII) detection, span extraction, and document de-identification.
The dataset contains 100,000 Urdu examples annotated with character-level spans identifying sensitive information such as personal names, addresses, phone numbers, email addresses, dates, and account/reference numbers.
The dataset is designed for training and evaluating Urdu PII detection and redaction models, including token-classification models, span-based NER systems, lightweight local LLMs, and privacy-preserving document processing pipelines.
| Property | Details |
|---|---|
| Language | Urdu |
| Task | PII Detection / Span Extraction |
| Format | JSONL |
| Total Records | 100,000 |
| Training Set | 80,000 |
| Validation Set | 10,000 |
| Test Set | 10,000 |
| Annotation Type | Character-level spans |
| Domains | Multiple real-world application domains |
The dataset contains synthetic Urdu examples curated to represent situations in which PII may appear in documents, messages, forms, records, and other text-based content.
The dataset is provided in JSONL (JSON Lines) format.
Each line represents one independent annotated example.
Each JSON object contains exactly three fields:
| Field | Type | Description |
|---|---|---|
text | string | Original Urdu document, sentence, or message. |
spans | object | Character-level PII annotations grouped by entity type. |
domain | string | Application domain represented by the example. |
The dataset is divided into three standard splits:
| Split | Records |
|---|---|
| Train | 80,000 |
| Validation | 10,000 |
| Test | 10,000 |
| Total | 100,000 |
The training split can be used for model training, the validation split for hyperparameter tuning and model selection, and the test split for final evaluation.
The dataset contains multiple categories of personally identifiable information.
| Label | Description |
|---|---|
private_person | Personal names and names of individuals |
private_address | Addresses, locations, institutions, branches, offices, and other location-related information |
private_phone | Personal telephone and mobile numbers |
private_email | Email addresses |
private_date | Dates associated with personal or sensitive records |
account_number | Account numbers, case numbers, policy numbers, reference IDs, claim IDs, document IDs, billing IDs, and similar identifiers |
A label can contain multiple spans when the same PII category occurs more than once in a document.
All PII annotations use character-level offsets.
Each span is represented as:
[start_offset, end_offset]
For example:
means that the characters between offset 24 and offset 35 correspond to a private_person entity.
Multiple entities of the same type are represented using multiple span pairs:
Similarly:
indicates two separate account/reference-number spans.
The following examples illustrate the annotation format and the types of PII represented in the dataset.
This example contains a person's name and identification/reference numbers in a government-related context.
This example contains a person's name, policy identifier, and email address.
This example demonstrates PII occurring in a social-media context, including a person's name, address, and identification number.
This example contains an email address, date, location information, and multiple reference identifiers.
This legal-domain example contains names, an identification number, a case number, location information, and a date.
This example demonstrates detection of personal names and a phone number in a news/media context.
This example contains institutional information, a billing identifier, an amount, and a date.
This example represents a travel-related claim containing identifiers, financial information, and a date.
This example demonstrates multiple PII categories occurring together, including identifiers, date, location, person's name, and phone number.
The dataset covers a variety of application domains where PII may occur.
Examples include:
legal
finance
healthcare
government
telecommunications
education
insurance
travel
ecommerce
news_channel
social_media
human_resources
The domain field allows models to be evaluated across different contextual environments.
The primary task supported by this dataset is Urdu PII span detection.
Given an Urdu text:
اینکر حسن رضا نے پروگرام میں بلال احمد کا فون 0335-7856037 ظاہر ہونے پر معذرت کی۔
A model should identify the PII spans:
The detected spans can then be passed to a downstream redaction system.