List your company

Safety Evaluation on HEx-PHI Alpaca risk-ranked subsets

13.1S1 ASR (%)Self-Inf-N
10.175629.915349.65569.3947May 6, 2026

Evaluation results

MethodLinks
Self-Inf-NBase Model=Qwen3-8B2026.0513.113.7930.3435.8652.41-39.31—
Self-Inf-NBase Model=Llama2-7b-Chat2026.0516.913.124.4837.5925.86-8.96—
Bi-Anchor(Grad)Base Model=Qwen3-8B2026.0517.932033.4523.4533.45-15.52—
Bi-Anchor(Reps)Base Model=Llama3.1-8B-Instruct2026.0518.6217.9330.6926.5559.66-41.04—
LARFBase Model=Llama2-7b-Chat2026.0519.3117.2419.3132.4125.17-5.86—
Bi-Anchor(Reps)Base Model=Llama2-7b-Chat2026.0526.5536.2121.3825.8634.14-7.59—
Bi-Anchor(Reps)Base Model=Qwen3-8B2026.0528.2837.2437.9327.5917.9310.35—
Bi-Anchor(Grad)Base Model=Llama2-7b-Chat2026.0529.6624.1418.9716.932.07-2.41—
Bi-Anchor(Grad)Base Model=Llama3.1-8B-Instruct2026.0537.2438.9724.8351.7235.521.72—
LARFBase Model=Qwen3-8B2026.054035.5219.3119.6616.923.1—
SQSD(Beaver)Base Model=Llama2-7b-Chat2026.0547.5924.8319.667.248.6238.97—
Reward ModelBase Model=Llama2-7b-Chat2026.0550.3416.916.97.599.3141.03—
SQSD(Aegis)Base Model=Qwen3-8B2026.0557.2435.5233.122.0714.8342.41—
SQSD(Aegis)Base Model=Llama2-7b-Chat2026.0557.5927.2419.3111.031047.59—
SQSD(Beaver)Base Model=Qwen3-8B2026.0561.3828.9727.5918.2811.0350.35—
Reward ModelBase Model=Qwen3-8B2026.0566.5517.5930.3421.7228.9737.58—
LARFBase Model=Llama3.1-8B-Instruct2026.0571.3859.3159.318.6220.3451.04—
Reward ModelBase Model=Llama3.1-8B-Instruct2026.0579.6615.5235.1717.5936.2143.45—
SQSD(Aegis)Base Model=Llama3.1-8B-Instruct2026.0581.3834.4823.126.912.7668.62—
SQSD(Beaver)Base Model=Llama3.1-8B-Instruct2026.0582.4127.9326.5510.691.3881.03—
Self-Inf-NBase Model=Llama3.1-8B-Instruct2026.0586.2136.930.3465.8660.3425.87—