aandreychuk/MAPF-GPT

该数据集包含10亿个观察-动作对，用于通过GPT模型克隆LaCAM方法来解决多智能体路径规划（MAPF）问题。数据集由Anton Andreychuk和Alexey Skrynnik策划，采用MIT许可证发布。数据集分为训练和验证两部分，训练部分包含1,000 * 2^20个观察-动作对，分为500个.arrow文件；验证部分包含2^20个观察-动作对，保存在一个.arrow文件中。整个数据集需要256GB的磁盘空间。更多关于数据集创建、源数据、观察结构等的详细信息可在相关论文中找到。

This dataset contains 1 billion observation-action pairs to clone the behavior of the LaCAM approach to solve the Multi-Agent Path Finding (MAPF) problem via a transformer (GPT) model. The dataset is divided into training and validation parts, with the training part containing 1,000 * 2^20 observation-action pairs divided into 500 .arrow files, and the validation part containing 2^20 observation-action pairs saved into a single .arrow file. The dataset requires 256 GB of disk space.