SafeMERGE addresses a critical challenge in LLM deployment: maintaining safety guardrails while fine-tuning models for specific applications. This novel approach selectively merges layers from safety-aligned and task-optimized models.

Key Innovations:

• Preserves safety alignment that typically deteriorates during standard fine-tuning
• Uses cosine similarity to identify and merge only layers that deviate from safe behavior
• Maintains task performance while significantly reducing harmful outputs
• Works as a post-fine-tuning technique, making it adaptable to various models

Medical Impact: SafeMERGE shows promising results on medical datasets like PubMedQA, enabling safer deployment of LLMs in healthcare settings where accuracy and safety are paramount.

This approach offers a practical solution for organizations seeking to customize AI capabilities while maintaining strong safety guardrails.

SafeMERGE: Preserving Safety Alignment in Fine-Tuned Large Language Models via Selective Layer-Wise Model Merging