The data used in my programs usually starts out as CSV files. If they are large and reading them takes a long time, I can read the data and then write it to a file using an unformatted stream. Reading such a file is very fast if you know the data types of the data written. How do you document this? The unformatted stream files I have created are only intelligible if I remember what programs wrote them, and that concerns me.
Maybe whenever you create an unformatted stream file foo.bin you should create a file foo.fmt describing the contents, for example