I am not aware of an algorithm to generate an optimal or even a good dictionary. This is generally done by hand. I think that a suffix tree would be a good approach to finding common strings for a dictionary, but I have never tried it.
The first thing to try is to simply concatenate 32K worth of your 1-3K examples and see how much gain that provides over no dictionary. Then you mess with it from there, changing the ordering of examples or pulling out repeated pieces in the examples to the end of the dictionary.
Note that the most common strings should be put at the end, since shorter distances take fewer bits.