PerplexityPrivacy models
PII-Tracer
PII-Tracer detects sensitive spans in conversational text.
What the release provides
Its approximately 600M bidirectional Qwen3-derived encoder has a 37-label token head covering nine private-information categories and a separate conversation-sensitivity head. Constrained decoding forms spans. The checkpoint uses MIT terms.PII-Tracer official model card
How to use it well
Practical guidance. Measure missed private spans and false matches independently. Retain the mapping from detected spans to original text when applying masking.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | perplexity-ai/PII-Tracer |