Piiranha: an open model for redacting personal information
A fine-tuned mDeBERTa token classifier that marks personal details in text across six languages, with more than a million downloads on Hugging Face.
Visit project ↗Hi nameJordan Avery, your order ships to address48 Lantern Row, cityBrightwater. Questions? Write emailorders@example.com.
Patient surnameR. Okonkwo (born birth date04/11/1987) can be reached at phone555-0147 before the follow-up.
Transfer confirmed from account account0000 1234 5678 for nameLior Mendes in cityAshford.
Before a dataset can be shared or a transcript published, the personal details in it have to come out. Piiranha is a token classification model that marks those spans so they can be redacted automatically. I trained it with the Internet Integrity Initiative, on GPUs sponsored by the Akash Network.
Details
- Base model. Microsoft’s mDeBERTa-v3-base, fine-tuned on the open ai4privacy masking dataset.
- Coverage. Seventeen categories of personal detail, in English, Spanish, French, German, Italian, and Dutch.
- Accuracy. 99.44% token accuracy overall, and 98.27% recall on the tokens that matter, measured on a held-out set of roughly 73,000 sentences.
The model has passed a million downloads on Hugging Face. There is a Colab quickstart if you want to try it on your own text.