List your company

Targeted Refusal on LLaMA2-13B-Chat (test)

100BadNetsNo Defense
12.79635.435558.07580.7145Mar 13, 2026

Evaluation results

MethodLinks
No DefenseThreat Model=N/A2026.0310074.8683.0796.5384.2884.75
CROWThreat Model=Full Params2026.0398.9832.5750.693.8766.1568.43
FTThreat Model=Lora Adapter2026.0398.5634.2652.3295.9468.3369.88
FTThreat Model=Full Params2026.0398.5475.6381.2697.2486.4587.82
PruningThreat Model=Lora Adapter2026.0398.3274.2181.2595.3787.2487.28
PruningThreat Model=Full Params2026.0393.846.7854.8695.8384.5275.16
QuantizationThreat Model=Full Params2026.0393.2135.6248.3784.878.6267.92
Fine-PruningThreat Model=Lora Adapter2026.0390.172.5483.4389.5278.4282.8
Fine-PruningThreat Model=Full Params2026.0383.6534.8662.7882.2545.3361.77
BD-VAXThreat Model=Full Params2026.0330.1624.3226.6432.3418.8626.46
BD-VAXThreat Model=Lora Adapter2026.0316.159.8312.6518.237.8212.94