If you have some idea of the correct structural alignment of your sequence family, you can provide a starting model of that alignment as a hint to hmmt. For instance, bashford.slx contains a structural alignment of 7 globin sequences.
> hmmb hint.hmm bashford.slx
> hmmt -v -i hint.hmm globin2.hmm globins50.fa
The -v option shuts off simulated annealing and substitutes a straightforward minimization (the Viterbi approximation to Baum-Welch expectation maximization). Alteratively, you can use the -B option, which shuts off annealing and substitutes full Baum-Welch expectation maximization. Full Baum-Welch seems better, but is less well tested and takes much longer.
It doesn't matter whether the sequences in the seed alignment file have counterparts in the training database. The initial alignment file is only used to generate the initial HMM, which is then used to align the sequences of the training database. The seed alignment does not contribute to the final statistics of the HMM. In fact, the final HMM will probably have a slightly different idea of how best to align your seed alignment.
If you can't be bothered to produce a careful manual alignment, and you're in too much of a hurry to do simulated annealing, I've found that I can get good alignments quickly by running an automatic alignment program such as CLUSTALV on some or all of the sequences, then using that alignment as a seed for hmmt. I think simulated annealing usually produces slightly better alignments, though.
Generally, starting from a structural alignment gives the best scores; simulated annealing may get close; simple minimization ( -v or -B options) is usually pretty bad. The score seems to be a good indication of the relative quality of the alignments that the model will produce; higher-scoring models give better alignments.