From MAME's -listxml output it's rather easy. Look at a set, check if there multiple identical crc32 values and compare the given sha1 values. If such values differ you got a set with collisions.
From a cmpro/frontend point of view you need to ensure that if you got crc32 based routines (for e.g. name checking), they need to take the sha1 values into account (at least for sets with collisions).