Hi,
We have a sum type of 151 constructors (and growing) in one project. There is a somewhat natural grouping of the constructors into 24 groups (with between 1 and 32 constructors in each group) so we used that to break it down into the two levels:
data Group1 = Ctor1 Int | Ctor2 Bool | ...
data Group2 = Ctor5 String | Ctor6 Double | ...
...
data A = Group1 Group1 | Group2 Group2 | ...
This wasn't enough, so then we cheated and wrote some code-gen. The datatype is described as data (think YAML or JSON) and then there's a short program which generates the declarations, including Haddock comments, and various useful functions such as somewhat-custom JSON serialisation. Each group gets its own module, which gives faster recompilation on changes. We could have used TemplateHaskell, except we wouldn't have got such nice Haddock docs (and, ew, TemplateHaskell) and it would all have had to have been in one module.
Hope that helps,
David