mahwizzzz / aegis

Aegis

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.

Domain Distribution (Train)

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

Per-Entity Overlap F1

Citation

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

@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.