← Evals

Evaluation profile

Claude 2 model-card safety and alignment evaluations

4sub-evals
0.152%total index weight
2components

Within-component eval weight: Truthfulness 0.607% · Misuse resistance 0.605%.

Model score (see each sub-eval direction)Predicted score

About this eval

Human-rated harmlessness and honesty, red-team robustness ordering, and combined HHH response recognition.

Included in the behavior ranking.

Sub-evals

MeasureComponentDirectionTotal index weightWithin-component weight
hhhclaude-system-cards/claude-2-scoring.csv:hhhMeasures whether a model recognizes the preferred response across combined HHH comparisons.truthfulness_honesty:0.500|ordinary_harm_misuse_resistance:0.500anthropic-claude-2-system-cardHigher is better0.0425%Truthfulness 0.202% · Misuse resistance 0.121%
human_feedback_harmless_eloclaude-system-cards/claude-2-scoring.csv:human_feedback_harmless_eloMeasures which model responses raters judged less harmful in adversarial conversations.ordinary_harm_misuse_resistance:1.000anthropic-claude-2-system-cardHigher is better0.0242%Misuse resistance 0.242%
human_feedback_honest_eloclaude-system-cards/claude-2-scoring.csv:human_feedback_honest_eloMeasures which model responses raters judged more accurate and factual.truthfulness_honesty:1.000anthropic-claude-2-system-cardHigher is better0.0607%Truthfulness 0.405%
red_teaming_rankclaude-system-cards/claude-2-scoring.csv:red_teaming_rankRanks the evaluated Claude systems by performance under Anthropic's red-team evaluation.ordinary_harm_misuse_resistance:1.000anthropic-claude-2-system-cardLower is better0.0242%Misuse resistance 0.242%

hhh

Measures whether a model recognizes the preferred response across combined HHH comparisons.

RankModelValueRelative performanceProvenance
1claude-20.86official
2claude-1.30.85official
3claude-instant-1.10.84official

human_feedback_harmless_elo

Measures which model responses raters judged less harmful in adversarial conversations.

RankModelValueRelative performanceProvenance
1claude-1.3282official
2claude-2278official
3claude-instant-1.1243official

human_feedback_honest_elo

Measures which model responses raters judged more accurate and factual.

RankModelValueRelative performanceProvenance
1claude-2137official
2claude-1.3102official
3claude-instant-1.168official

red_teaming_rank

Ranks the evaluated Claude systems by performance under Anthropic's red-team evaluation.

RankModelValueRelative performanceProvenance
1claude-21official
2claude-1.32official
3claude-instant-1.13official