AdvBench
Benchmarks
Task NameDataset NameSOTA ResultTrendResults
advbench
3.8Raw Jailbreak Success Rate
3
AdvBench Adv (test)
43.5MS_R
3
AdvBench (full)
100Good Bot Rate
3
AdvBench
94Refusal Rate (AdvBench)
3
AdvBench Multiple-choice format (full)
99Safe Option Probability
3
AdvBench PAD (test)
14ASR (GPT-4o)
3
AdvBench Vicuna to gpt-3.5-turbo (transfer)
90ASR (Refusal Matching)
3
AdvBench
0.95Win Rate
3
AdvBench
64ASR (PAIR)
2
AdvBench Malicious-Educator
50Attack Success Rate (ASR)
2
AdvBench RPO defense (full)
92ASR
2
AdvBench PAT defense (full)
80ASR
2
AdvBench First-Step GCG (test)
4ASR (GPT-4o)
2
AdvBench DiJA (test)
18ASR (GPT-4o)
2
AdvBench Llama to gpt-3.5-turbo (transfer)
87ASR (Refusal Matching)
2
AdvBench (held-out)
15Union Match Rate
1
AdvBench (subset)
30Fires Rate
1
AdvBench
99.2AUROC (vs Alpaca)
1
AdvBench 520 prompts
93Accuracy
1
AdvBench
52.23PPL
1
AdvBench
37.6ASR@1
1
AdvBench
13.27Success Rate @1
1