SVG benchmarking is a thing since GPT-4, so probably all major labs are overfitting on some dataset ov svg images for sure