Safety Evaluation Set
Benchmarks
TaskDatasetSOTA resultTrendResults
Safety Evaluation Set
1.66Harmful Response Rate
25
Safety Evaluation Set GCG, AutoDAN, DeepInception, Prefilling, Intent Laundering n=200 (full public set)
0GCG Attack Success Rate
11
Safety Evaluation Set Moderation (held-out target labels)
0.89AUROC
6
Safety Evaluation Set Sentiment (held-out target labels)
97.5AUROC
6
Safety Evaluation Set Jailbreaking (held-out target labels)
97.4AUROC
6
Safety Evaluation Set Toxicity (held-out target labels)
97.6AUROC
6