Hi Andre, you should first optimize your data processing as much as possible ... try different programs, and do test refinements against the data they give you. Unfortunately Rwork/Rfree is a poor indicator unless the Wilson B is the same; what is much more robust is the correlation coefficient CCwork/CCfree as a function of resolution.
If you get 25% Rfree in twin refinement in a low-symmetry spacegroup with equal twin fractions, and 33% in the high-symmetry spacegroup without twin refinement, then the high-symmetry spacegroup may be the correct one. As Phil says, the intensity stats may be skewed due to close spots. To answer your question: for an unbiased free set, assign the free reflections in the high-symmetry space group and expand them (using sftools or similar) to the low symmetry space group. best, Kay
