List your company

Machine Unlearning on WMDP-bio 1.0 (test)

80.3BF16 AccuracyZero-shot
20.39635.94851.567.052May 14, 2026

Evaluation results

MethodLinks
Zero-shotModel=Qwen-3-8B2026.0580.380.30—74.1
Zero-shotModel=Qwen-3-4B2026.0573.173.10—65.7
SimNPOModel=Qwen-3-4B2026.0572.773.40.00774.163.1
GU+SimNPOModel=Qwen-3-4B2026.0572.472.70.00373.863.1
NPOModel=Qwen-3-4B2026.0571.870.70.01173.162.4
NPOModel=Qwen-2.5-4B2026.0570.7690.01871.358.4
SimNPOModel=Qwen-2.5-4B2026.0570.369.70.00771.559.1
GU+SimNPOModel=Qwen-2.5-4B2026.0569.8700.00271.559.1
Zero-shotModel=Qwen-2.5-4B2026.0569.369.30—61
Zero-shotModel=Llama-3.2-3B2026.0568.768.70—53.8
SimNPOModel=Llama-3.2-3B2026.0566.567.50.016450.5
GU+SimNPOModel=Llama-3.2-3B2026.0566.266.80.00764.351.4
NPOModel=Llama-3.2-3B2026.0564.863.20.01765.250
MANSUModel=Qwen-3-8B2026.0561.758.10.03667.172.9
Surgical GAModel=Llama-3.1-8B2026.0554.757.30.0275648.3
MANSUModel=Qwen-3-4B2026.0552.7490.03761.764.4
MANSUModel=Qwen-2.5-4B2026.0548.745.40.03358.859.8
NPOModel=Llama-3.1-8B2026.0544.342.30.0250.356.3
MANSUModel=Llama-3.2-3B2026.0543.440.40.0352.852.1
MANSUModel=Llama-3.1-8B2026.0543390.0452.357.3
NPOModel=Qwen-3-8B2026.0528.3320.03730.349.2
GU+SimNPOModel=Qwen-3-8B2026.0526.726.30.00327.756.8
Global GAModel=Llama-3.1-8B2026.0526310.052623.5
Surgical GAModel=Qwen-3-8B2026.052624.70.01330.345.8
SimNPOModel=Llama-3.1-8B2026.0524.324.70.00330.355.3
Global GAModel=Qwen-3-8B2026.0523.324.30.00724.724.2
GU+SimNPOModel=Llama-3.1-8B2026.052323024.720
SimNPOModel=Qwen-3-8B2026.0522.722.7025.726.5