In the end a benchmark does nothing useful. So every benchmark has some artificial constraints that give it a purpose. I think this benchmark is meant to have string-like keys. Perhaps because it's supposed to work on other input files too that have a column of alphanumeric fields. So for every benchmark there's a need for a person that takes partially arbitrary decisions regarding where to draw a line between a fair solution and an unfair one. I think a fair solution of this problem has string-like keys. Your solution is based on a different assumption. Both are equally valid until someone defines what's fair for this benchmark.
In my third version I've also kept the code short because I think it's more reasonable this way 
But your solution is interesting and fast.
This seems to accept solutions in other languages too: