My preference is to validate external data and turn into a “known” data structure at the edges, and let it crash once you know you have good data. I think this applies a bit more to the case where you have truly external data such as console input or tcp data but also to your public API of your library.
I personally don’t care to much about the error message from a library. I.e :badarg is OK as I then can look up the documentation.
If you are going to “let it crash” or not on bad input depends on the consequences of the crash and if it is only going to happen during development or if it can happen in production. It is used quite a bit in OTP libraries so I guess it OK to do so 
