LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B
This work highlights a practical vulnerability in AI safety training, showing that fine-tuning can efficiently undo safety measures, which is an incremental but critical issue for AI developers and risk assessors.
The researchers tackled the problem of safety alignment robustness in large language models by subversively fine-tuning Llama 2-Chat models, achieving refusal rates of about 1% on harmful instructions while retaining general performance.
AI developers often apply safety alignment procedures to prevent the misuse of their AI systems. For example, before Meta released Llama 2-Chat - a collection of instruction fine-tuned large language models - they invested heavily in safety training, incorporating extensive red-teaming and reinforcement learning from human feedback. We explore the robustness of safety training in language models by subversively fine-tuning Llama 2-Chat. We employ quantized low-rank adaptation (LoRA) as an efficient fine-tuning method. With a budget of less than \$200 and using only one GPU, we successfully undo the safety training of Llama 2-Chat models of sizes 7B, 13B, and 70B and on the Mixtral instruct model. Specifically, our fine-tuning technique significantly reduces the rate at which the model refuses to follow harmful instructions. We achieve refusal rates of about 1\% for our 70B Llama 2-Chat model on two refusal benchmarks. Simultaneously, our method retains capabilities across two general performance benchmarks. We show that subversive fine-tuning is practical and effective, and hence argue that evaluating risks from fine-tuning should be a core part of risk assessments for releasing model weights. While there is considerable uncertainty about the scope of risks from current models, future models will have significantly more dangerous capabilities.