somukandula/maskara-indian-pii-200k

Maskara Indian PII Dataset (Phase 2)

Synthetic Indian PII dataset for training the Maskara NER model. This version has been expanded and diversified for Phase 2 training.

Splits

Split	Rows	Purpose
train	~250,000	Model training
template_disjoint_eval	15,000	Generalization evaluation: entire template families held out
real_world_eval	~2,600	Manually curated real-world Indian text evaluation

Schema

Column	Type	Description
id	string	Unique example identifier
text	string	Raw text containing PII
entities	list[dict]	Character-level spans: {start, end, label}
domain	string	Source domain (e.g. whatsapp, email, json, ocr)
locale	string	Text locale (en-IN, en-US, hi-IN)
difficulty	string	easy, medium, hard, hard_negative
generator	string	Generator version (maskara-v2)
seed	int	Per-example random seed
template_family	string	Template family for traceability

Supported Entities (17)

ADDRESS, API_KEY, CREDIT_CARD, DATE_OF_BIRTH, DRIVER_LICENSE, EMAIL, IP_ADDRESS, PASSWORD, PERSON_NAME, PHONE, SSN, USERNAME, AADHAAR, PAN_CARD, PASSPORT, UPI_ID, VEHICLE_REG.

Note: LOCATION was removed in Phase 2.

Diversity Improvements

• Templates: 150+ multi-entity templates, 80+ single-entity templates.
• Formatting variations: 6–10 per entity (plain, spaced, dashed, labelled, etc.).
• Hinglish / code-mixed: ~40% of examples.
• Hard negatives: ~15% of examples.
• Structured / noisy documents: ~20% of examples.
• Domains: WhatsApp, SMS, email, support tickets, HTML, Markdown, JSON, CSV, SQL, logs, OCR, bank statements, government forms, Aadhaar letters, PAN documents, resumes, configuration files.

Template-Disjoint Evaluation

The following template families are excluded from training and reserved for the template_disjoint_eval split:

• passport_application
• transport_challan
• json_payload
• config_yaml
• html_user
• api_request_json
• sms_link
• finance_pipe
• identity_kyc
• medical_record

Real-World Evaluation

A manually curated split containing realistic Indian text such as:

• Customer support conversations
• Hinglish chats
• OCR-noisy documents
• Government forms
• JSON payloads
• API request bodies
• Log lines
• Bank statements
• Emails
• Transport challans

This split is never used during training.

Usage

from datasets import load_dataset

dataset = load_dataset("somukandula/maskara-indian-pii-200k")
print(dataset["train"][0])

Generation

Generated using maskara_dataset_generator.py (Maskara Phase 2 generator). Each example is produced by filling placeholder-based templates with synthetic values and recording character-level entity spans.

Phase 2 Error-Analysis Augmentation

After the first Phase 2 training run, targeted synthetic examples were added for entities that underperformed on real-world and template-disjoint evaluation:

• CREDIT_CARD
• PASSWORD
• DRIVER_LICENSE
• USERNAME
• ADDRESS
• IP_ADDRESS
• PHONE
• AADHAAR
• PAN_CARD

These examples live in phase2/targeted_augmentation.py and are merged into the train split before retraining.

Files in this repository

File	Description
maskara_dataset_generator.py	Full synthetic dataset generator
generate_maskara_phase2_dataset.py	Script that generates and pushes all splits
phase2/targeted_augmentation.py	Targeted augmentation for weak entities
phase2/train_maskara.py	Training pipeline
phase2/evaluate_maskara.py	Evaluation pipeline
phase2/error_analysis_loop.py	Error analysis and actionable report generator
phase2/retrain_with_augmentation.py	End-to-end retraining with augmentation

Citation

@misc{maskara_dataset2026,
  title={Maskara Indian PII Dataset},
  author={Somu Kandula},
  year={2026},
  howpublished={\url{https://huggingface.co/datasets/somukandula/maskara-indian-pii-200k}}
}

Generated by ML Intern

This dataset repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

• Try ML Intern: https://smolagents-ml-intern.hf.space
• Source code: https://github.com/huggingface/ml-intern