ResearchBenchmarksAlignment defense against harmful fine-tuning on GSM8KFollow0.91Harmful Score (HS)Antibody-0.58329.495919.57529.6541Feb 28, 2026Evaluation ResultsMethodMethodLinksHarmful Score (HS)False Acceptance Rate (FA)AntibodyBackbone=Gemma-2-9BBackbone=Gemma-2-9B2026.020.9157.43LisaBackbone=Gemma-2-9BBackbone=Gemma-2-9B2026.0213.0254.97BoosterBackbone=Gemma-2-9BBackbone=Gemma-2-9B2026.0222.1358.97SFTBackbone=Gemma-2-9BBackbone=Gemma-2-9B2026.0232.0556.73VaccineBackbone=Gemma-2-9BBackbone=Gemma-2-9B2026.0238.2454.37