The obvious first suggestion would be to also carry out the CCSD extrapolation using smaller basis sets, i.e. W[TQ,TQ,DTQ].
The effect for the SCF component of TAE is very small on condition that the extrapolation is carried out not on the individual total energies but rather on the computed SCF components of TAE themselves. Clearly error compensation occurs between the molecule and the constituent atoms.
The effect on the valence correlation component, unfortunately, is rather more significant. Over the `training set', MAE rises to 0.37 kcal/mol even after SO2 (which clearly is a pathological case here) has been eliminated. Aside from the latter, eliminating systems with mild nondynamical correlation does not lead to any significant reduction in MAE. Also noteworthy is that, on average, the binding energies appear to be somewhat overestimated: this is easily explained from the fact that basis sets like AVTZ+2d1f are not quite saturated in the radial correlation energy either, and that therefore the TAE[AVQZ+2d1f]-TAE[AVTZ+2d1f] gap will be an overestimate of the TAE[L=4]-TAE[L=3] gap.