Incomple/DuET-PD
收藏官方服务：
资源简介：
DuET-PD是一个综合性的框架和用于评估大型语言模型在多轮说服性对话中的鲁棒性和适应性的数据集。它包含2246个多项选择题，每个问题都附带了一系列预生成的说服性呼吁，旨在挑战LLM在三个对话回合中的初始立场。数据集来源于MMLU-Pro和SALAD-Bench，分别涉及知识密集型场景和安全关键上下文。
DuET-PD is a comprehensive framework and dataset designed to evaluate the robustness and adaptability of Large Language Models (LLMs) in multi-turn persuasive dialogues. It contains 2,246 multiple-choice questions, each augmented with a series of pre-generated persuasive appeals challenging an LLMs initial stance over three conversational turns. The dataset sources from MMLU-Pro and SALAD-Bench, covering knowledge-intensive scenarios and safety-critical contexts respectively.
提供机构：
Incomple
