linzju/Code-Llama-3-8B_EnchTable_FFN

EnchTable: Unified Safety Alignment Transfer in Fine-tuned LLMs

This repository contains the Code-Llama-3-8B model aligned using the FFN (Feed-Forward Network) variant 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.

Model Details

• Name: Code-Llama-3-8B (EnchTable-FFN)
• Base Model: Llama-3-8B (Fine-tuned for Code)
• Method: EnchTable (FFN Module)
• Primary Use Case: Safety Alignment Transfer / Secure Code Generation

EnchTable is a novel framework designed to transfer safety alignment capabilities from a safe source model to various fine-tuned target models (e.g., Domain-Specific LLMs) without compromising their downstream performance.

This specific checkpoint represents the FFN-based intervention, where safety vectors are calculated and merged specifically into the Feed-Forward Network layers of the model to mitigate harmful outputs while preserving code generation capabilities.