gSCAN

gSCAN数据集由阿姆斯特丹大学创建，专注于评估机器在基于语言理解的场景中的组合泛化能力。该数据集包含超过30万条指令及其对应的行动序列，涉及多种组合泛化类型，如颜色和形状的新组合、新方向的导航等。数据集通过将语言指令与网格世界的状态相结合，使模型能够学习基于语言规则的泛化。gSCAN的应用领域包括提升机器学习模型在自然语言处理中的泛化能力和效率，以及增强模型在复杂环境中的适应性和灵活性。

The gSCAN dataset was created by the University of Amsterdam, focusing on evaluating the compositional generalization ability of machines in language understanding-based scenarios. It contains over 300,000 instruction-action sequence pairs, covering multiple types of compositional generalization such as novel combinations of colors and shapes, navigation in new directions, and so on. By combining linguistic instructions with the states of grid worlds, the dataset enables models to learn generalization based on language rules. The application fields of gSCAN include improving the generalization ability and efficiency of machine learning models in natural language processing, as well as enhancing the adaptability and flexibility of models in complex environments.