Dataset for: Sycophancy Under Social Pressure in Medical LLMs: A Dose-Response Study Across Models and Languages

This dataset accompanies the study "Sycophancy Under Social Pressure in Medical LLMs: A Dose-Response Study Across Models and Languages" (El Husseini, 2026). The dataset contains 6,300 challenge-response exchanges from 7 large language models (GPT-4o, GPT-4o-mini, Claude Sonnet 4, Qwen2.5-32B, Qwen2.5-14B-Instruct, Gemma3-12B, Ministral-14B) tested on 300 medical multiple-choice questions in English (MedQA-USMLE) and French (MediQAl) under three levels of social challenge intensity (light, moderate, strong). Contents: Question text Question-level metadata (language, cognitive classification, source dataset, question ID) Model responses (initial and revised) for all 6,300 exchanges Extracted single-letter answers and grading outcomes Transformer-derived psycholinguistic scores (sentiment, agreement, certainty, uncertainty, politeness, assertiveness) for initial and revised responses本数据集配套于研究"《医疗大语言模型（Large Language Model，LLM）在社会压力下的谄媚行为：跨模型与语言的剂量反应研究》"（El Husseini, 2026）。 本数据集包含来自7个大语言模型（GPT-4o、GPT-4o-mini、Claude Sonnet 4、Qwen2.5-32B、Qwen2.5-14B-Instruct、Gemma3-12B、Ministral-14B）的6300轮挑战-应答交互数据。前述模型基于300道医学多项选择题接受测试，题目分别源自英语数据集MedQA-USMLE与法语数据集MediQAl，并设置了三级社会挑战强度（轻度、中度、重度）。 数据集内容包括： 问题文本 问题级元数据（语言、认知分类、源数据集、问题ID） 全部6300轮交互的模型应答（初始应答与修正后应答） 提取的单字母答案与评分结果 针对初始与修正后应答，由Transformer衍生的心理语言学评分（涵盖情感、契合度、确定性、不确定性、礼貌程度、自信程度）