To be fair, once it does generalize the pattern then the benchmark is actually measuring something useful for deciding if the model will be able to product a subject-verb-object SVG.