The drift time predictor training is not yet available from the command-line, but I added a TODO issue for this, which you can track (https://skyline.ms/issues/home/issues/details.view?issueId=679).
The predictor training takes the spectrum for the apex time of the selected peak and slides the extraction window (defined by resolving power on Agilent) and slides it across the precursor m/z window calculating the signal it would extract at each IMS value. It then takes the IMS value with the maximum extracted signal. This avoids having to do peak detection and dealing with multiple conformers, etc.
Hope this is helpful. Sorry, we haven't exposed training at the command line yet.