PrivacyFilterTuned

PrivacyFilterTuned is a finetuned version of openai/privacy-filter. It is adapted for privacy filtering workflows where partial SSNs and related compact identifiers should be captured more consistently.

The main target behavior is improved detection of account-number style spans such as:

• SSN ending 1234
• SSN last four are 1234
• partial SSN 1234
• nearby MRN, member ID, student ID, and similar identifier patterns

The model keeps the original Privacy Filter label space. Partial SSNs are labeled as account_number.

Base Model

This checkpoint is based on openai/privacy-filter, a bidirectional token classification model for PII detection and masking. The base model detects the following span categories:

• account_number
• private_address
• private_date
• private_email
• private_person
• private_phone
• private_url
• secret

Finetuning

This checkpoint was finetuned locally on a small targeted dataset of healthcare, education, and finance-style prompts. The data focuses on partial SSNs and identifier boundary cases.

Limitations

This model is a privacy-filtering aid, not an anonymization or compliance guarantee. It may still miss sensitive spans or redact spans that should remain visible. Evaluate it on your own in-domain data before production use, and keep human review paths for high-sensitivity workflows.

The finetuning set is small and targeted, so improvements are expected mainly around partial SSNs and similar compact identifiers rather than broad PII coverage.

Usage

With this repository downloaded locally:

python -m opf "Patient Alice Smith has SSN ending 1234." \
  --checkpoint /path/to/privacyfiltertuned \
  --device cpu

For evaluation:

python -m opf eval /path/to/eval.jsonl \
  --checkpoint /path/to/privacyfiltertuned \
  --device cpu