it is actually demonstrated that The straightforward pre-instruction activity of predicting which caption goes with which image can be an efficient and scalable way to understand SOTA graphic representations from https://k2spiceshop.com/product/liquid-k2-on-paper-online/