EnchTable: Unified Safety Alignment Transfer (Bio-Medical + Attention)
This repository contains the Bio-Medical Llama-3-8B model aligned using the Attention-based intervention of the EnchTable framework.
This model is part of the research presented in the paper: "EnchTable: Unified Safety Alignment Transfer in Fine-tuned Large Language Models", accepted at IEEE S&P 2026.
Base Model: Llama-3-8B fine-tuned on medical datasets
EnchTable is a framework designed to transfer safety alignment capabilities from a general-purpose safe model to specialized downstream models.
In this checkpoint, we apply safety vectors specifically to the Self-Attention layers. This ensures the model reduces harmful outputs (hallucinations, toxicity) while maintaining high performance in medical question-answering tasks.
