Welcome! Type "help" for available commands.
$
Summary
A 600M-parameter bidirectional Qwen3 encoder model from Perplexity AI that masks 9 PII categories (names, emails, phones, addresses, dates, URLs, account numbers, other PII, secrets) in conversational text.
It uses a BIOES token classification head with a constrained Viterbi decoder for span extraction, plus a sensitivity head for conversation-level scoring. Max sequence length is 4096 tokens; inference requires torch, safetensors, and transformers.