It’s empirical. The most important values are how much computer time you have and how much you want out of the results.
One model is enough to tell you if your inputs are correct.
A thousand models is enough to tell you if Rosetta can produce good structures for your model – if all 1000 are obviously bad, then either there’s an error in the input or our search algorithms aren’t up to the task.
Production runs are generally 10,000 to a million models. Broadly, you can stop modeling when your models produce a “funnel”, where the all the low-energy structures are similar to each other, and as structures become less similar to the low-energy structure their scores also get worse. I think if you have a great template you’ll be on the low end of this range.
.
