Pay more attention to the robustness of LLMs on adversarial prompt for instruction data mining.

Abstract

Instruction tuning has emerged as a paramount method for tailoring the behaviors of LLMs. Recent studies have unveiled the potential for LLMs to achieve high performance through fine-tuning with a limited quantity of high-quality instruction data. Instruction-Following Difficulty is one of the most representative approaches in instruction data mining, which involves selecting samples where LLMs fail to generate response that align with the provided instructions as the high-quality instruction data. Building upon this approach, we further investigate how the robustness of LLMs to adversarial prompts influences the selection of high-quality instruction data. This paper proposes a pioneering framework of high-quality instruction data mining for instruction tuning, focusing on the impact of LLMs' robustness on adversarial prompts. Our notable innovation is to generate adversarial instruction data by attacking the prompts associated with instruction samples. Then, we introduce an Adversarial Instruction-Following Difficulty (AIFD) metric, which utilizes complete instruction sample pairs to identify samples with high adversarial instruction difficulty as high-quality instruction data. Apart from it, to address cases where LLM responses deviate from user intent, we further introduce a novel Adversarial Instruction Output Embedding Consistency (AIOEC) method that relies solely on instruction prompts to mine high-quality online instruction data. We conduct extensive experiments on two benchmark datasets to assess the performance. The experimental results serve to underscore the effectiveness of our proposed two methods. Moreover, the results underscore the critical practical significance of considering the robustness of LLMs on adversarial prompts for instruction data mining.