If that was the case, the models would have been producing near perfect outputs for it a year ago.
Instead they are just training on general SVG generation, which in no way should be viewed as "benchmaxxing".
