Ironically, at work we buy a bunch of BLS datasets from a third-party vendor. Their format is even worse: an opaque binary database and a win32 DLL that reads it. 1000 lines of Haskell later, it sort of works...

(This is mostly because they chose to represent dates as integers, and they have four possible date types: yearly, quarterly, monthly, and daily. 1901, in their exciting world, could mean "year 1901" or it could mean "month January 1919". That's nice work, Lou.)