How to deal with multiple splits (for example, 2 train splits, 3 test splits)

[pfliu-nlp]

#246 (comment)

[neubig]

For test sets, I think that just calling them what they're called in the original paper might be a good idea. For example dev devtest test?

For training sets, that's an interesting question. If the different training datasets correspond to different training settings then maybe they should be different subdatasets (like we did for fig_qa).