OPEN_SOURCE ↗
REDDIT · REDDIT// 5h agoMODEL RELEASE
OpenAI ships local PII filter
OpenAI released Privacy Filter, an Apache 2.0 open-weight model for detecting and redacting personally identifiable information in text. The 1.5B-parameter token classifier runs locally, supports 128K-token context, and targets training, indexing, logging, and review pipelines.
// ANALYSIS
This is a practical infrastructure release, not a flashy frontier model, and that makes it more useful for developers handling sensitive corpora.
- –Local redaction matters because raw PII can stay on-device before data enters model training, RAG, analytics, or support workflows
- –OpenAI claims 96% F1 on PII-Masking-300k, rising to 97.43% on its corrected benchmark, but teams still need in-domain evaluation
- –The fixed taxonomy covers names, addresses, emails, phones, URLs, dates, account numbers, and secrets, so custom policies will require tuning
- –Apache 2.0 licensing and Hugging Face/GitHub availability make this unusually deployable for commercial privacy tooling
// TAGS
openai-privacy-filteropen-weightsself-hosteddata-toolssafetyllm
DISCOVERED
5h ago
2026-04-22
PUBLISHED
5h ago
2026-04-22
RELEVANCE
8/ 10
AUTHOR
ai_hedge_fund