How much of this could be resolved if its training set were reduced? Conceivably, most of the training serves only to confuse the model when only aiming to solve a math equation.