readme should just be straightforward, provide different pretrained model weights.
I can create a benchmark table with columns: model name, parameters/size, modality, testing metrics etc... and have a fixed seed for this benchmark..
Need to figure out hosting of these models first