Note that the .mol2 file must have the correct bond order for the C=N bond, or the CGenFF server may be confused. The H atom placement in the molecule, esp. protonation states, must also be correct.
A high penalty doesn't necessarily mean the estimated parameters are wrong, just that the analogy to existing validated parameters is not very good. It means at least testing the estimated parameters, which should be done for whatever approach is used to get a first test set, CGenFF or chemical intuition. Multiple parameter sets are often tested in simulations, so it is not unreasonable to test both choices and the CGenFF estimates.
It should be noted that QM calculations are encouraged, if possible, when the penalties are very large.
The problem for testing is often the difficulty in finding good target data from experiments that help refine the conformational preferences in the appropriate environment, e.g. water and/or a protein binding site. Aggressive and persistent searching of the literature and other sources for the whole molecule or relevant fragments is often needed to find some useful data.
If the molecule will form an important part of the research project, the decisions made on these parameters are critical. Some additional, unplanned effort may be required in order to ensure producing a respectable result for the parameters.