PII Detection Models (pt-BR / LGPD)

/ Overview

Most redaction tooling is built for English and trained on the kinds of identifiers US compliance cares about. Brazilian Portuguese, and the LGPD’s specific categories of sensitive data, fall through the gaps. This project is a set of open models, two live demos, and a benchmark aimed squarely at that gap: one demo uses GLiNER with flexible labels, and the other uses a fine-tune of OpenAI’s Privacy Filter.

The point is practical privacy: catch the data that actually has to be removed before a document leaves the building, in the language and legal frame it was written in.

/ What it does

• Tags personal and sensitive data in pt-BR text at the token level, including the categories LGPD treats as sensitive.
• Ships as open models with a runnable demo, so the behaviour can be inspected rather than trusted on faith.
• Publishes the predictions and metrics behind the comparison, so the trade-offs are visible, not asserted.

/ What the benchmark showed

Two model families are compared: a fine-tune of OpenAI’s Privacy Filter, and smaller GLiNER models. Neither wins everywhere. On the in-distribution Portuguese validation set they are within a point of each other (partial F1 0.897 against 0.888). On a held-out Portuguese source the GLiNER model is well ahead (0.900 against 0.704). On text that contains no personal data at all the order reverses: across 2,419 spam and phishing messages, the Privacy Filter fine-tune raised 64 false positives and the GLiNER model raised 8,666.

For a redaction tool that last number matters as much as F1: a model that flags names everywhere produces documents nobody can read, and reviewers who stop trusting the flags.

/ Approach

Evaluated in the open, with error reporting ahead of a headline number. All evaluation data so far is synthetic; a human-annotated, real-world test set and comparisons against off-the-shelf tools are the next steps, and the results here should be read with that in mind.