GRASP
收藏资源简介:
GRASP数据集由信息科学研究所 USC Viterbi 工程学院创建,是一个用于评估大型语言模型(LLMs)在网格环境中常识空间推理能力的大型基准。该数据集包含16000个不同的网格实例,每个实例都是一个二维数组,包含空单元格、障碍物或能量单元。数据集的创建过程涉及多种能量分布模式和障碍配置,旨在模拟现实世界中的导航和资源收集场景。GRASP数据集主要用于评估和提升LLMs在空间推理和规划方面的能力,特别是在复杂环境中的导航和资源管理。
The GRASP dataset was created by the Information Sciences Institute of the USC Viterbi School of Engineering, and it is a large-scale benchmark for evaluating the commonsense spatial reasoning capabilities of Large Language Models (LLMs) in grid-based environments. This dataset contains 16,000 distinct grid instances, each of which is a two-dimensional array composed of empty cells, obstacles, or energy units. The creation of this dataset involves multiple energy distribution patterns and obstacle configurations, aiming to simulate real-world navigation and resource collection scenarios. The GRASP dataset is primarily used to evaluate and enhance the spatial reasoning and planning capabilities of LLMs, particularly for navigation and resource management in complex environments.




