The quality of classical biomolecular simulations is inevitably limited by two problems: the accuracy of the force field used and the comprehensiveness of configuration space sampling. In this work we tackle the sampling problem by carrying out driven adiabatic free energy dynamics to obtain converged free energy surfaces of dipeptides in the gas phase and in solution using selected dihedral angles as collective variables. To calculate populations of conformational macrostates observed in experiment, we introduce a fuzzy clustering algorithm in collective-variable space, which delineates macrostates without prior definition of arbitrary boundaries. With this approach, we calculate the conformational preferences of small peptides with six biomolecular force fields chosen from among the most recent and widely used. We assess the accuracy of each force field against recently published Raman or IR-UV spectroscopy measurements of conformer populations for the dipeptides in solution or in the gas phase.