Why it matters

Toxicity filter shapes UX + safety. Understanding shapes moderation.

Advertisement

The architecture

Classifier scores per category (identity attack + insult + threat + severe toxicity).

Threshold + action.

Toxicity filter flowScore categoriesidentity + insult + threat + etc.Thresholdper categoryActionblock / warn / logPerspective API + Detoxify + Llama Guard; layer for coverage
Toxicity filter.
Advertisement

How it works end to end

Perspective API (Jigsaw).

Detoxify open-source.

Llama Guard.

Category-specific tuning.