A privacy-filter model for detecting and redacting personally identifiable information (PII) in Urdu text across legal, financial, healthcare, government, and other domains. The model is fine-tuned from openai/privacy-filter on mahwizzzz/uopf-sipar, a 100K Urdu dataset spanning twelve domains.

The model performs span level detection of names, phone numbers, email addresses, addresses, dates, and account/CNIC-style identifiers in Urdu text.

Training Dataset

The model was trained on mahwizzzz/uopf-sipar, a Urdu dataset for OPF style privacy span detection.

Split	Records
Train	80,000
Validation	10,000
Test	10,000

Domain Distribution (Train)

Domain	Records
legal	21,986
finance	13,850
telecom	6,315
healthcare	5,410
travel	4,876
social_media	4,845
news_channel	4,818
government	3,965
ecommerce	3,923
insurance	3,382
education	3,319
hr_jobs	3,311

Evaluation

The model was evaluated on a held out test split with character level PII spans. The primary comparison is between the original base model, openai/privacy-filter, and this fine-tuned checkpoint.

The headline metric is same label overlap span F1. This metric counts a prediction as correct when it overlaps the gold span and has the correct entity label. It is more appropriate for redaction-oriented systems than exact character matching, because small boundary differences usually do not prevent a sensitive region from being detected.

Exact span F1 is also reported. It requires the predicted label, start offset, and end offset to match the gold annotation exactly.

Base vs Fine-tuned Model

Metric	Base openai/privacy-filter	Fine-tuned model	Absolute Gain
Overlap Detection F1	0.517	0.957	+0.440
Overlap Precision	0.497	0.997	+0.500
Overlap Recall	0.538	0.920	+0.382
Macro F1	0.468	0.931	+0.463
Exact Span F1	0.021	0.406	+0.385

Per-Entity Overlap F1

Entity	Base F1	Fine-tuned F1	Absolute Gain
private_person	0.509	0.966	+0.457
private_phone	0.386	0.755	+0.369
private_email	0.646	0.999	+0.353
private_address	0.015	0.986	+0.971
private_date	0.657	0.989	+0.332
account_number	0.597	0.893	+0.296

Citation

If you use this model in research or production, please cite:

bibtex

@misc{mahwiz2026urdu_privacy_filter,
  title        = {Aegis},
  author       = {Mahwiz Khalil},
  year         = {2026},
  howpublished = {Hugging Face Model Hub},
  url          = {https://huggingface.co/mahwizzzz/aegis}
}

License

Apache 2.0. See LICENSE for details.