Hi, Alasdair.
I find that the number of articles contained in "nytimes-2020-04-21.gz" does not agree with the number reported in your paper.
In Table 2 of "Transform and tell", the number of training, validation, and test splits are 433561, 2978, and 8375, but the MongoDB backup file you provided contains 434314, 3052, and 8495 articles.
Did I do something wrong? Or why did the number of articles in the NYTimes dataset grow?