Jan Lisec pointed me to a severe bug in the functional group prediction feature implemented in the GMD. It turned out that I normalised spectra before decision tree training different than spectra for spectral classification. Jan pointed me to a spectrum where the Phosphoric Acid Deriv group was predicted present based on m/z 299 although this particular mass had only a minimal intensity in this GC-MS spectrum.
The only good news is that the validation for the publication was not affected by this bug, because the cross validation is performed without this web interface. However, I removed this bug and want to apologise for this inconvenience.
the other Jan ;-)