Models are collected by running a slightly modified version of CPMpy, which stores (pickles) the models that are given to a solver.
(this covers both the case of a generic model.solve() call, as well as of a specific model.solve("mysolver") call)
See patch included.
We also compressed them post-hoc with 'brotli' to save diskspace, code not included...
Also some helper scripts to detect unsat models, and to identify too slow models for inclusion in the testsuite, as well as removing structurally likely-similar models.