Safety guardrail for text

Detects negative safety classes — aggression, threats, forbidden requests, spam and more — and highlights the spans the model would redact. Joint sequence + token-level head trained on a curated safety corpus.