SORRY-Bench
Benchmarks
Task NameDataset NameSOTA ResultTrendResults
Sorry-Bench
99.09Safety Score
90
SORRY-BENCH
0.9Alignment Faking Rate
46
SORRY-Bench
10.22ASR
40
SORRY-Bench
0.9Expected HFR
35
Sorry-Bench base
92.73Safety Score
27
SORRY-Bench
13ASR
24
SORRY-Bench audio
78.41ASR Accuracy
15
SORRY-Bench
85.7Score
14
SORRY-Bench
26.9Compliance Rate
12
SORRY-Bench
70.45Refusal Rate
7
Sorry-Bench (test)
93.9Accepted HF Accuracy
6
Sorry-Bench
85.5Refusal Rate (String Match)
6
Sorry-Bench
65.7Ref %
6
SORRY-Bench linguistic mutations
25.4Compliance Rate
6
SORRY-Bench decontaminated base prompts
17.1Compliance Rate
6