I believe the 79% comes from this reportFermion ReportAddendum
However, in the report the group compared the 3 different samples in pairs. So the algorithms they used only tried to figure out which one of two samples matched. So the comparisons were for A vs B, A vs C and B vs C. Also, the 79% is from the best pair. It may also be worth noting that Fermion had 1500 samples like the contest but used 1200 samples for training and then tested the remaining 300 samples.
If my math is right you could run all 3 Fermion algorithms against each sample and get them to "vote" on which of the 3 samples it is likely to be. In this case you get about 47% accuracy. So I think the actual accuracy achieved is 47% not the 79%.