Dear Luc,
The basic idea for forcing to have a minimum number of sites is that one can't infer a good tree from too few sites (imagine trying to build a tree on 5000 sequences from 50 sites). Hence, GARD demands at least p + 2 sites per partition, where p is the number of parameters (number of tree branches + global parameters).
For your second dataset (49 species), there are 2*49-3 = 95 branch lengths PER partition. GARD will expect at least two, so there are 190 parameters to estimate for branch lengths alone. As you can see, we are trying to estimate almost as many parameters as there are observations (sites), which is statistically untenable.
You can run single breakpoint partition analysis on any dataset (e.g. using Multimedia File Viewing and Clickable Links are available for Registered Members only!! You need to
) and base your inference on AIC (not c-AIC), which does not demand a minimum number of sites.
Alternatively, reduce the size of your alignment by pruning sequences. You can remove those that are closely related (e.g. pick one per clade), but this is very heuristic and should be dealt with based on what your data are.
Cheers,
Sergei