Protegrity releases an open model that finds 69 types of personal data

Protegrity, a company for data protection, launched Protegrity-PII, a model that finds personal data in text. With open weights, a person can download the model and operate it. The model has 278 million parameters, the numbers that set how a model works. It sorts 69 types of personal data, such as card data, health data, credentials and biometric data.

How it works

The model marks each token, a small piece of text, as personal data or not. It treats text as content to check and not as instructions to follow. Attacks that put instructions in text have less effect on it. A firm can operate it in its systems.

The numbers

Recall is the share of personal data that the model finds. Precision is the share of its flags that are correct. Protegrity reports that the model processed one million characters in 0.49 seconds, for about $0.0005, on one high end graphics chip.

What is not known

Protegrity made the test, and the results are from Protegrity. The source gives no test by a different party. The authors give the method in a report on arXiv.

Insight: The test maker also tops the test

Protegrity wrote PII-SCOPE and is first on it. A firm can operate the model on its hardware, as it can with Muse Spark Open from Meta. This seems to be a bid for the standard in personal data tools, because the firm that sets the test sets the score. Test the model on your data.