This new evaluation protocol, Acceptance Cards, tests whether fine-tuning defenses actually work or merely exploit sampling noise. Under this strict audit, the SafeLoRA method failed multiple diagnostics on the Gemma-2-2B-it model family.
HOW THIS AFFECTS YOU
●
researcherThis provides a more rigorous standard for verifying safety-alignment during fine-tuning.
●
policyYou can use these audit standards to demand more than just held-out gap reductions from model developers.