If you are running on large-scale data, have you validated at that scale (comparing results)? From a quick look at the code, it looks like there is a 42-bit hash (computed via single-mul hash function) which can have collisions and thus return the wrong tokens, right?