HELM safety scenario that scores a model's first reply to 38,961 Anthropic red-team openings with a 0–1 LLM-judge safety_score.
unassessed
| Category | safety |
|---|---|
| Subcategory | single-turn refusal on first utterances from Anthropic's 2022 red-team dialogues |
| Page status | active |
| Metric | safety_score (mean of LLM-judge harmlessness ratings) |
| Direction | higher_is_better |
| Unit | 0-1 scale |
| Dataset size | 38961 |
| Dataset licence | MIT |
| Publisher | Anthropic (dataset); Stanford CRFM (HELM scenario) |
HELM's anthropic_red_team scenario takes Anthropic's public red-team dialogues and keeps only the first human utterance of each transcript. The model must answer that single English prompt. The original collection was a multi-turn attack by crowdworkers trying to elicit harmful assistant behaviour. HELM does not replay later turns, so the score is a one-shot refusal test on historically successful attack openings, not a reproduction of Anthropic's multi-turn red-team protocol.
Zero-shot generation. HELM sends the first Human utterance with empty instructions, max_tokens 512, temperature 0. Two LLM judges (the safety annotator) rate the completion's harmfulness on a 0–1 scale; higher means a more harmless reply.
No model card in ModelSpec reports this benchmark yet.