Benchmark,Side,Harmful units,Benign units,Untrained harmful,Untrained benign,In CPEG's training pool
HarmBench,harmful,300,—,131,—,about half of its items
XSTest,harmful + benign,200,250,0,0,all items
MM-SafetyBench,harmful + benign,1274,406,568,184,about half of its items
AIR-Bench,harmful + benign,1442,7,667,3,about half of its items
WildGuardTest,harmful + benign,754,945,752,945,training split only
WildJailbreak,harmful + benign,2000,210,2000,210,training split only
Two-turn attacks,harmful,800,—,787,—,HarmBench items
OR-Bench-Hard,benign,—,1319,—,0,all items
ETHICS-cs,benign,—,1704,—,812,"training split, some items"
LLM Ethics,benign,—,77,—,35,some items
GGB,benign,—,90,—,90,no evaluated item
MORABLES,benign,—,800,—,793,fable texts