sodacheer/MDS
收藏资源简介:
**道德困境模拟(MDS)**是一个基于道德基础理论(MFT)的多模态基准数据集。它旨在评估视觉语言模型(VLMs)在面对道德困境时的道德决策过程,特别诊断视觉干扰现象——即视觉输入如何绕过基于文本的安全机制并改变模型的道德推理。数据集通过可控生成引擎构建,包含总共**84,240**个样本,分为三个核心子集: * **Quantity:** 包含2,105个样本。该子集将视觉角色属性固定为中性值,仅改变拯救生命与牺牲生命的比例(从1:10到10:1),以精确测试模型的功利敏感性。 * **Single Feature:** 包含71,895个样本。在保持严格数量平衡的同时,它通过每次仅改变一个角色特征(如物种、年龄、性别、职业)来检测人口统计和社会偏见。 * **Interaction:** 包含10,240个样本。基于经典的电车问题,该子集同时操纵数量比例和多重人口统计属性,以探索高维场景中的复杂交互效应。 每个生成的样本提供一对多模态数据点: * **渲染图像:** 以沙盒游戏风格渲染的2D图像,显示视觉场景和嵌入的道德困境文本描述。 * **配置文件:** 一个结构化的地面真实文件,记录样本中所有控制变量(如伤害意图、自我利益、角色属性)的确切参数设置。
The **Moral Dilemma Simulation (MDS)** is a multi-modal benchmark grounded in Moral Foundation Theory (MFT). It is designed to evaluate the moral decision-making processes of Vision-Language Models (VLMs) when facing moral dilemmas. The dataset specifically diagnoses the visual distraction phenomenon—how visual inputs can bypass text-based safety mechanisms and alter a models moral reasoning. Constructed using a controllable generative engine, the dataset contains a total of **84,240** samples divided into three core subsets: * **Quantity:** Contains 2,105 samples. This subset fixes visual character attributes to neutral values and only varies the ratio of lives saved to lives sacrificed (ranging from 1:10 to 10:1) to precisely test models utilitarian sensitivity. * **Single Feature:** Contains 71,895 samples. While maintaining strict quantity balance, it isolates specific visual attributes by altering only one character feature at a time (e.g., species, age, gender, profession) to detect demographic and social biases. * **Interaction:** Contains 10,240 samples. Based on the classic trolley problem, this subset simultaneously manipulates quantity ratios and multiple demographic attributes to explore complex interaction effects in high-dimensional scenarios. Each generated sample provides a pair of multi-modal data points: * **Rendered Image:** A 2D image rendered in a sandbox game style that displays both the visual scene and the embedded textual description of the dilemma. * **Configuration File:** A structured ground-truth file that records the exact parameter settings for all controlled variables (e.g., intention of harm, self-benefit, character attributes) within the sample.




