Since you seem to already have a good idea of what you want to have happen, I would suggest writing some tests that demonstrate that your function behaves as you wish it to.

Often times I find that, by the time I have the tests written enough to be useful, the code almost writes itself.. and it makes the really hard things possible by keeping a hard copy around of what getting each little piece working entails.

If you have a representative sample of the data you are trying to reformat, it might help to pull out some 'hard' cases from it, or indeed start with a simplistic solution, and use that until you can find some data that it breaks with.