These are json triples of dicts for (topic, doc, and labels). There are preprocessing method in ctmatch that need to be brought into a tokenizer maybe, that I use to make the formats needed by the models used, but this is a start.
Also, it's private until I confirm that both the TREC and KZ data are already on here, in the ir_datasets dataset... .