ResearchBenchmarksAlignment Defense against Harmful Fine-tuning on SST2Follow11.3Harmful Score (HS)Antibody9.68620.580531.47542.3695Feb 28, 2026Evaluation ResultsMethodMethodLinksHarmful Score (HS)False Acceptance Rate (FA)AntibodyBackbone=Gemma-2-9BBackbone=Gemma-2-9B2026.0211.394.27BoosterBackbone=Gemma-2-9BBackbone=Gemma-2-9B2026.0228.4794.15LisaBackbone=Gemma-2-9BBackbone=Gemma-2-9B2026.0234.6794.04SFTBackbone=Gemma-2-9BBackbone=Gemma-2-9B2026.0240.7794.11VaccineBackbone=Gemma-2-9BBackbone=Gemma-2-9B2026.0251.6594.5