TXUXIv3

该数据集是对AIXI-Shape数据集的扩展，包含了52,000张简单图形的图片，这些图片的背景选自Describable Textures Dataset，具有非均匀性，从而使得生成更多域外样本成为可能。该数据集的目标是模拟真实数据集的局限性，允许背景具有更多样性。规模上，该数据集包含52,000张图片，其任务是用决策树解释来评估保真度指标。

This dataset is an extension of the AIXI-Shape dataset, containing 52,000 images of simple geometric shapes. The backgrounds of these images are sourced from the Describable Textures Dataset, which are non-uniform, thereby enabling the creation of additional out-of-distribution samples. The core objective of this dataset is to simulate the limitations of real-world datasets, allowing for greater diversity in background content. With a total of 52,000 images, this dataset is utilized to evaluate fidelity metrics through decision tree explanations.