List your company

Machine Unlearning on WMDP-Cyber (Pre/Post Accuracy)

41.5Accuracy (Pre-Unlearning)Original model
23.92428.48733.0537.613Apr 15, 2026

Evaluation results

MethodLinks
Original model2026.0441.5—
SimNPOretain free=false, Time=545s2026.0440.841.4
RMUretain free=false, Time=1420s2026.0440.239.8
NPOretain free=false, Time=1178s2026.0439.641
MC-WIN-Uretain free=true, Time=50679s2026.0437.639.5
GradDiffretain free=false, Time=280s2026.0427.437.4
GradAscentretain free=true, Time=173s2026.0424.626.5