Discovered while working on #935: np.float64 arrays serialized using the svmlight formatter loose precision beyond single precision.