Bio-Medical Llama-3-8B with EnchTable Safety Alignment
This model, linzju/Bio-Medical-Llama-3-8B_EnchTable_FFN, is an 8 billion parameter variant of Llama-3, specifically fine-tuned for the bio-medical and healthcare domain. It incorporates a unique safety alignment methodology from the EnchTable framework, as detailed in a paper accepted at IEEE S&P 2026.
Key Capabilities & Differentiators
- Domain-Specific Knowledge: Built upon a Llama-3-8B base model that has been fine-tuned on medical datasets, ensuring strong performance in bio-medical tasks.
- Unified Safety Alignment: Employs the EnchTable framework to transfer safety alignment capabilities, a method designed to make specialized models safer.
- FFN-based Intervention: Uniquely applies safety vectors specifically to the Feed-Forward Network (FFN) layers. This targeted approach is designed to effectively mitigate safety risks without compromising the model's factual knowledge essential for bio-medical applications.
- Balanced Performance: Aims to maintain high factual accuracy in its specialized domain while enhancing safety, addressing a critical need for LLMs in sensitive fields like healthcare.
Ideal Use Cases
- Applications requiring a balance of specialized bio-medical knowledge and enhanced safety.
- Research and development of safe AI tools in healthcare.
- Tasks where mitigating potential safety risks in LLM outputs is paramount, particularly in medical contexts.