erg0dic/STaR

系统化、多路径、分离的时空推理（STaR）基准数据集。该数据集的主要思想是将二进制关系组合的概念扩展为不仅原子性，还要求模型能够通过多条路径而不是单一路径进行推理，与之前的艺术作品相比具有显著区别。数据集的复杂性有两个维度：路径长度k（源节点和目标节点之间的路径长度）和路径数量b（源节点和目标节点之间的路径数量）。每个数据集都包含边索引、边标签、查询边和查询标签等信息。训练集包含小的图，路径长度为2、3、4，路径数量为1、2、3。其余的数据集用于测试系统性泛化，目标类别是平衡的。

Systematic, Multipath, disjunctive Spatio-Temporal Reasoning (STaR) benchmark. The main idea is to expand the concept of binary relational composition to not just be atomic but also require that a model reason over multiple paths instead of a single one, in contrast with previous art. The complexity of the dataset has two dimensions: path length k (the length of the path between the source and target nodes) and the number of paths b (the number of paths between the source and target nodes). Each dataset contains edge index, edge labels, query edge, and query label information. The training sets contain small graphs with path lengths of 2, 3, 4, and the number of paths of 1, 2, 3. The rest of the datasets are for testing systematic generalization, and the target classes are balanced.