The high-redshift universe remains one of the last unexplored realms of astrophysics and the standard model of cosmology. Measurement of the highly-redshifted 21-cm emission line of neutral hydrogen allows for a direct probe of both the Dark Ages before the first luminous astrophysical objects, and the Cosmic Dawn when those first sources appeared. Contamination, however, from ubiquitous, bright foregrounds and coupling of the spectral structure between the latter and a beam through chromaticity can obviate statistically significant signal extraction, especially as accurate foreground models rely on precise maps of the low-frequency radio sky, of which there is a dearth. To overcome these difficulties, a beam-weighted foreground model which incorporates chromaticity as well as inherent uncertainty in the radio sky maps must be used. We present preliminary results from a suite of simulated signal extractions involving realistic foregrounds, beams, and signals. We show how robust training sets for the foreground are generated by increasing the spectral and spatial uncertainty in the latter and thus “widening” the foreground model. Even in the case where the true foreground model is significantly uncertain, as is the case for all realistic 21-cm global signal experiments, robust signal extraction through such a training set-based pipeline can be achieved.