The dark signal is exposure time dependent. The one appropriate for a given observation has to be derived by interpolation between dark frames of different exposure times. Note that long dark exposures first have to be cleaned of particle hits (see above). Bias and dark signal will usually be determined and subtracted in the same step.
In a CCD with good cosmetic properties, bias and dark signal are the same for all pixels. Then, only a number (obtained from, e.g., STATISTICS/IMAGE) should be subtracted in order to prevent the noise in the dark frames to propagate into the reduced spectra. If, because of local variations, this is not possible, some suitable smoothing should still be attempted.