OpenAI Releases Privacy Filter: 1.5B Parameter PII Desensitization Model Open-Sourced

OpenAI recently released Privacy Filter, a model for desensitizing personal identifiable information (PII) in text. It is open-sourced under the Apache 2.0 license on Hugging Face and GitHub, allowing developers to download, customize, and deploy it commercially.

The model uses a bidirectional Token classification architecture with a mixed expert (MoE) design—total 1.5 billion parameters, but only about 50 million are active at a time. It can run on laptops or even browsers, supporting a context window of up to 128,000 tokens.

It can identify 8 types of sensitive information such as names and addresses. After fixing dataset annotation issues, its F1 score reached 97.43%. Moreover, it processes data locally to avoid the risk of sensitive information exposure.