OpenAI released Privacy Filter, an open-weight model under the Apache 2.0 license that detects and redacts PII in text and is available on Hugging Face and GitHub. The small token-classification model runs locally, supports long contexts, classifies eight PII categories, and achieved about 96–97% F1 on the PII-Masking-300k benchmark, though OpenAI advises human review for high-sensitivity use cases. #PrivacyFilter #OpenAI
Keypoints
- Privacy Filter detects and redacts PII in unstructured text while distinguishing public from private information.
- The model is open-weight under the Apache 2.0 license and available on Hugging Face and GitHub.
- It runs locally using a token-classification approach, supports up to 128,000-token contexts, and uses 1.5B parameters with ~50M active at runtime.
- Privacy Filter categorizes sensitive data into eight groups, including names, addresses, emails, phone numbers, URLs, dates, account numbers, and secrets.
- OpenAI reports ~96–97% F1 on the PII-Masking-300k benchmark but notes potential misses and recommends domain-specific fine-tuning and human review.