Why is AM so small??

[ShellingFord221]

Hi, in the paper, there are 1,666,764 entities and 5,988,321 edges reported in the dataset of AM

However, in rgcn/data/am, there is only 1000 instances randomly sampled from the complete AM dataset:

I wonder that results of AM reported in the paper are based on the complete dataset or just the sampled set? I think this may cause very huge performance margin against reported results:

There is also no sampling process illustrated in the original paper:

Besides the results, I also wonder that whether the model can run on a large-scale dataset. Thanks!