YinYangAlign
收藏资源简介:
YinYangAlign是由南卡罗来纳大学人工智能研究所和Meta AI共同创建的一个高级基准测试框架,该框架旨在量化文本到图像系统(T2I)的校准保真度,涵盖六个根本性和固有的矛盾设计目标。数据集包含详细的人类提示、选择的响应、被拒绝的AI生成输出以及对底层矛盾的说明。该数据集用于评估和优化T2I系统在多个互相矛盾的目标之间的平衡能力。
YinYangAlign is an advanced benchmarking framework co-developed by the University of South Carolina Artificial Intelligence Institute and Meta AI. This framework is designed to quantify the calibration fidelity of text-to-image (T2I) systems, covering six fundamental and inherent contradictory design objectives. The dataset contains detailed human-written prompts, curated responses, rejected AI-generated outputs, and explanations of the underlying contradictions. It is utilized to evaluate and optimize the capability of T2I systems to strike a balance among multiple mutually conflicting goals.




