GradShield: Alignment Preserving Finetuning

GradShield is a filtering method that safeguards LLM alignment during finetuning by computing Finetuning Implicit Harmfulness Scores to identify and remove harmful data points before they corrupt model behavior.