We use the combined GALFORM semi-analytical model of galaxy formation and GRASIL spectrophotometric code to investigate the properties of galaxies selected via their sub-mm emission. Our fiducial model has previously been shown to fit the properties of local ULIRGs, as well as the number counts of faint sub-mm galaxies. Here, we test the model in detail by comparing the SEDs and stellar, dynamical, gas and halo masses of sub-mm galaxies against observational data. We precisely mimic the sub-mm and radio selection function of the observations and show that the predicted far-infrared properties of model galaxies with S_850>5mJy and S_1.4>30uJy are in good agreement with observations. Although the dust emission model does not assume a single dust temperature, the far-infrared SEDs are well described by single component modified black-body spectrum with characteristic temperature 32+/-5K. We also find evidence that the observations may have uncovered evolution in the far-infrared--radio relation in ULIRGs out to z~2. We show that the predicted redshift distribution of sub-mm galaxies provides a reasonable fit to the observational data with a median redshift z=2.0, with the radio-selected subset predicted to make up approximately 75% of the population. However, the predicted K-band and mid-infrared (3--8um) flux densities of the sub-mm galaxies (and LBGs) are up to a factor 10x fainter than observed. This discrepancy may indicate that the stellar masses of the sub-mm galaxies in the model are too low: M~10^10Mo, while observations suggest more massive systems, M~10^11Mo. Finally, we discuss the potential modifications to the models which may improve the fit to the observational data. [Abridged]