CRL HarmBench

Gemma

Baseline

• keyword based 35.36%

• rejection detection model based 44.64%

21th

• default 45.36%

• with mask generated 47.50%

• with mask all 45.71%

past

Corrsteer

• non-decode global 33.57% , decode global 34.29%

• 67.50% max pooling global

• 47.86% mean pooling global

• coeff mean, corr max: 48.21%

• coeff max, corr mean: 0%

LLama