When you train the model, use the option --random-seed INTEGER
(different from 0, otherwise it uses the clock) to fix the random seed. It should give you consistent results over multiple runs.
There was a bug with this feature, which is now fixed in the development release.
See MALLET's download page to build the most current version.