A Study on Prompt Types for Harmlessness Assessment of Large-Scale Language Models

This paper presents a study on prompt types for the harmlessness assessment of large-scale language models. In recent years, large-scale language models have revolutionized the field of natural language processing, and harmlessness assessment of the output of language models has become crucial. In this paper, we propose three types of prompts that can elicit harmful answers from large-scale language models. The generated prompts are then entered into publicly available language models (e.g., Llama-2-70b, GPT-3.5-Turbo-Instruct, Claude-instant-100k) to examine the results of eliciting harmful answers. The results provide useful guidance for improving model trustworthiness through harmlessness assessment of large-scale language models.