Recently, OpenAI released a new model called Privacy Filter, designed to help developers effectively anonymize personally identifiable information (PII) in text. The model has a parameter scale of 150 million and uses a mixture of experts (MoE) design. It is open-sourced under the Apache 2.0 license on Hugging Face and GitHub platforms, allowing developers to download, customize, and use it for commercial purposes.
The core advantage of Privacy Filter lies in its deep language understanding capabilities, enabling it to identify sensitive information in unstructured text through context. Unlike traditional rule-based privacy filtering tools, this model can accurately retain public information while masking or anonymizing sensitive data related to specific individuals. This capability allows developers to build more robust privacy protection mechanisms in training pipelines, indexing processes, log recording, and review stages.
The model supports a context window of up to 128,000 tokens and decodes coherent segments using a restricted Viterbi algorithm. In evaluations, Privacy Filter performed exceptionally well on the PII-Masking-300k benchmark test, achieving an F1 score of 96%. After correcting annotation issues identified during evaluation, the model's F1 score further increased to 97.43%, demonstrating its efficiency in identifying personal sensitive information.
OpenAI stated that Privacy Filter is not an anonymization tool and cannot replace compliance certification. In high-sensitivity scenarios such as legal, medical, and financial fields, manual review, and domain-specific assessment and fine-tuning remain essential. Additionally, Privacy Filter is designed to protect user privacy and can run on local devices, so users do not need to worry about personal information leakage when using AI tools.
