GSCAN

该数据集名为gSCAN，是一个评估情境语言理解中组合性泛化能力的基准，主要研究代理如何在网格世界环境中解释和执行命令。该数据集包含了训练和测试集，要求进行不同形式的语言泛化，包括对新颖的颜色-形状组合进行零样本泛化。数据集规模为19,282个测试示例，其任务旨在评估在接地语言理解中的系统性泛化能力。

This dataset, named gSCAN, is a benchmark for evaluating compositional generalization in situated language understanding. It primarily investigates how agents interpret and execute commands in grid-world environments. The dataset comprises training and test splits, with tasks requiring diverse forms of language generalization including zero-shot generalization to novel color-shape combinations. It contains 19,282 test examples, and its tasks aim to evaluate systematic generalization in grounded language understanding.