.svg)
What is Toxicity Metric?
A toxicity metric scores whether language-model output contains offensive, abusive, hateful, or otherwise harmful language so teams can fail builds that ship unsafe replies.
Use calibrated thresholds, multilingual samples, and adversarial phrasings; pair with refusal checks so models neither over-refuse benign queries nor leak toxic content.
Related Giskard articles
Catch toxic outputs with Giskard
Run harmful-content probes and guardrail evals so toxicity regressions surface before release. See Giskard Guards.
Further reading
Authoritative reference: Perspective API (toxicity research).
