Dear Karin,
RELY was originally designed to check reliability of coding tiers, rather than the main line. For that, it does a great job, because the shape of coding tiers is so well defined. If that is what you are doing, then I would say that the decision about what percentage to check is really a function of your overall corpus. For example, if you only have two hours of recording and you check 10% of that, then that would seem incomplete. However, if you have 200 hours and you check a random 10% of that and show good reliability, then I would be pretty convinced that you are okay. So, percentage is a function of corpus size.
I suppose a similar argument could be made regarding corpus or data type, but that argument would be a bit less convincing.
Later on, Leonid expanded the function of RELY to include comparisons of the main line. This method just looks for overall match of words. Please just read the manual about these differences.
Best,
-- Brian MacWhinney