This study focuses on Reynolds-averaged Navier Stokes (RANS) models for passive scalar transport. A recently developed machine learning turbulence closure, the tensor basis neural network for scalar flux modeling (TBNN-s), is applied on two different datasets, an inclined jet in crossflow and a wall mounted cube in crossflow. The TBNN-s consists of a deep neural network that predicts a turbulent diffusivity matrix after being trained on high fidelity datasets. The present work addresses the generality question (i.e. will a model trained in one turbulent flow perform well in a different turbulent flow?) by training and testing across different classes of flows. We show that when trained and tested in the same class of flow, the TBNN-s obtains significant improvements in the mean scalar field over a common baseline model. When trained and tested across different classes of flows, the TBNN-s has a similar performance to the baseline model, which is encouraging since training set extrapolation does not produce the nonphysical results some fear. Further analysis suggests that varying training sets for the same test case produces similar results in terms of alignment of the scalar flux vector, but different turbulent diffusivity matrix magnitudes.