alignmment

This paper proposes DataRx, a missingness-aware sampling method that selects safety-critical examples to preserve LLM safety during task-specific fine-tuning, reducing attack success rates from 59.23% to 13.70% on Llama3-8B-Instruct.