Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Safety Benchmarks

Benchmarks

Task NameDataset NameSOTA ResultTrend
Safety AlignmentSafety Benchmarks (Sorry-bench, StrongREJECT, WildJailbreak, JBB-PAIR, JBB-GCG)
Average Score42.34
21
Safety EvaluationSafety Benchmarks ASB, InjecAgent, AHarm, ASecBench
ASB Score61.1
18
Safety EvaluationSafety Benchmarks VLG-1, VLG-2, Ch3EF
VLG-1 Score93.91
17
Lifelong Safety AdaptationSafety Benchmarks (Final-day)
Final-day Macro F1 Score96.2
15
Text-to-Image Safety EvaluationSafety Benchmarks (MMA, Sneaky, I2P-S, Ring, P4D)
MMA80.4
10
Safety EvaluationSafety Benchmarks
Average ASR0.4
10
Jailbreak Attack EvaluationFive Safety Benchmarks AdvBench, HarmBench, HarmfulQ, JBBench, StrongReject
ASR7.69
6
Safety EvaluationSafety Benchmarks Overall
Cost per Accuracy Point ($)0.001
4
Safety EvaluationSafety Benchmarks Aggregate (test)
Generation Quality (Std Prefix)73.6
4
Safety EvaluationFive Safety Benchmarks direct_q
ASR0.02
3
Showing 10 of 10 rows