You can safely delete, rather than compress, those large databases.
For MD, there's really no reason to save the fingerprints as the odds of encountering an atom in the exact same relative position is almost zero. The databases are really for saving duplicate calculations in training, and we should ideally make an option to not save database files during force calls. We have this on our to-do list, but I'm not sure when it's going to rise to the top.
For MD, I'd strongly recommend trying the experimental LAMMPS interface put together by Efrem Braun and Levi Lentz. For now, you can find instructions on this mailing list -- there is an email from April 13 from Efrem that describes it. In our experience, it takes a couple of hours to get it all set up correctly, but once you do you can run MD much faster, parallelize the force calls, and it won't write any database files.