Several additional features also aid the learning performance of the SAM system. First, SAM performs the first training pass (before any surgery occurs) on multiple models. The best of those models is selected for the remaining surgery and training iterations. Second, SAM also supports Viterbi training, in which the model is trained according to the best path of each sequence through model, rather than by the probability distributions over all possible paths. Although this is slightly faster, the results are generally not as good. Third, when models are derived from an alignment or an existing profile, training of a module's match table or transitions can be turned off and it can be insulated from the surgery procedure. This last feature was not used for the results of this paper, apart from the related protection of FIMs from surgery.