OpenAI releases an open-weight, locally runnable model under Apache 2.0 for detecting and redacting PII in text with claimed state-of-the-art accuracy.
Key Takeaways
Available now on Hugging Face and GitHub under Apache 2.0; designed to run locally without calling OpenAI APIs.
Architecture: bidirectional token-classification model, labels a full input in one pass, then decodes spans via constrained Viterbi procedure.
1.5B total parameters, 50M effective parameters at inference – unusually lightweight for the task.
Current V2 release covers 8 PII entity types; starts from an autoregressive pretrained checkpoint adapted into a token classifier.
Hacker News Comment Review
Code inspection by a PII-systems builder found V4 and V7 taxonomies already in the repo with 20+ entity types – the 8-type V2 release looks like a staged rollout, not the full scope.
Commenters challenged the demos: most examples are simple name/email/phone patterns catchable by regex; the real test is ambiguous or context-dependent PII that regex fails on.
Prior art exists: SuperagentLM shipped on-edge PII redaction models at 200M, 3B, and 20B sizes years ago, giving practitioners a direct comparison baseline.
Notable Comments
@aubinkure: Found V4/V7 config taxonomies in source code with 20+ entity types, suggesting phased entity expansion is already planned.
@Havoc: Flags that 50M effective params is impressively light and asks whether a similarly sized prompt-injection detection model exists.