“So, if any of the tie-break labels is equal to the test label, it is marked as correct.”
The paper that had Twitter cheering gzip over neural nets used kNN with k=2 and scored a tie as correct if either neighbor matched. That is top-2 accuracy reported as plain accuracy. Corrected, KinyarwandaNews drops from 0.891 to 0.835 and SwahiliNews from 0.927 to 0.850. On KirundiNews gzip goes from best to worst, which is the Table 5 nobody tweeted.