List your company
ResearchFormerly WizWand

Harmful score evaluation on BeaverTails (test)

28.4Harmful ScoreSFT
-0.9286.68614.321.914Feb 5, 2026Feb 21, 2026Mar 9, 2026Mar 26, 2026Apr 11, 2026Apr 27, 2026May 14, 2026
Updated 5mo ago

Evaluation results

MethodLinks
SFTn=2000, Harmful Ratio (p)=0.12026.0228.4
SFTHarmful Ratio (p)=0.2, n=10002026.0228
SFT2026.0227.6
SFTHarmful Ratio (p)=0.1, n=10002026.0227.6
SFTn=Average, Harmful Ratio (p)=0.12026.0225.84
SFTHarmful Ratio (p)=0.15, n=10002026.0225.8
SFTn=1500, Harmful Ratio (p)=0.12026.0224.7
SFTn=500, Harmful Ratio (p)=0.12026.0224.4
SFTn=2500, Harmful Ratio (p)=0.12026.0224.1
SFTHarmful Ratio (p)=Average, n=10002026.0223.1
SFTHarmful Ratio (p)=0.05, n=10002026.0221
SPARD2026.0220.6
SPARDHarmful Ratio (p)=Average, n=10002026.0219.84
SafeGrad2026.0218.5
ConstrainedSFTHarmful Ratio (p)=Average, n=10002026.0217.8
ConstrainedSFT2026.0217.6
DSSHarmful Ratio (p)=Average, n=10002026.0216.86
DSS2026.0216.8
SafeGradHarmful Ratio (p)=Average, n=10002026.0216.3
AsFT2026.0216.1
AsFTHarmful Ratio (p)=Average, n=10002026.0215.96
SFTHarmful Ratio (p)=clean, n=10002026.0215.1
Lisa2026.0214.8
LisaHarmful Ratio (p)=Average, n=10002026.0213.88
Surgeryn=500, Harmful Ratio (p)=0.12026.029.7
SurgeryHarmful Ratio (p)=0.15, n=10002026.029.6
Surgeryn=1500, Harmful Ratio (p)=0.12026.029.4
SurgeryHarmful Ratio (p)=0.2, n=10002026.029.3
Surgeryn=Average, Harmful Ratio (p)=0.12026.028.94
Surgery2026.028.9
SurgeryHarmful Ratio (p)=0.1, n=10002026.028.9
SurgeryHarmful Ratio (p)=0.05, n=10002026.028.8
SurgeryHarmful Ratio (p)=Average, n=10002026.028.42
Surgeryn=2500, Harmful Ratio (p)=0.12026.028.4
Surgeryn=2000, Harmful Ratio (p)=0.12026.028.3
SurgeryHarmful Ratio (p)=clean, n=10002026.025.5
Unlearn-SmoothModel=OPTML-Group/NPO-SAM-WMDP, Attack=Adaptive attack2026.050.738
VAAModel=Llama-2-7B, Attack=Adaptive attack2026.050.707
VaccineModel=Llama-3.1-8B, Attack=Adaptive attack2026.050.707
VAAModel=Llama-3.1-8B, Attack=Adaptive attack2026.050.705
SDDModel=Llama-2-7B, Attack=Adaptive attack2026.050.705
VaccineModel=Llama-2-7B, Attack=Adaptive attack2026.050.701
SDDModel=Llama-3.1-8B, Attack=Adaptive attack2026.050.693
SDDModel=Qwen2-7B, Attack=Adaptive attack2026.050.691
VaccineModel=Qwen2-7B, Attack=Adaptive attack2026.050.5
BoosterModel=Llama-3.1-8B, Attack=Adaptive attack2026.050.416
CTRAPModel=Llama-3.1-8B, Attack=Adaptive attack2026.050.375
CTRAPModel=Qwen2-7B, Attack=Adaptive attack2026.050.324
BoosterModel=Qwen2-7B, Attack=Adaptive attack2026.050.307
VAAModel=Qwen2-7B, Attack=Adaptive attack2026.050.239
CTRAPModel=Llama-2-7B, Attack=Adaptive attack2026.050.229
BoosterModel=Llama-2-7B, Attack=Adaptive attack2026.050.2