REBUS

该数据集是一套用于评估模型在识别传达秘密代码单词的绘画方面的表现的对抗性数据集。在此数据集中，模型的性能表现出明显的差异，尤其在交叉熵损失与顶级准确度（Top-1 accuracy）之间存在显著分歧。所面临的任务是：在对抗性条件下进行绘画识别。

This is an adversarial dataset developed to evaluate model performance on the task of recognizing paintings that convey secret code words. This dataset shows significant discrepancies in model performance, particularly a striking divergence between Cross-Entropy Loss and Top-1 accuracy. The core task addressed by this dataset is painting recognition under adversarial conditions.