So a couple things surprised me… It turns out alot of these widely used models actually don’t publish their scaling behavior or do a really hand wavy bare min pass at it. And, scaling behavior is actually pretty different depending on the domain and dataset. Aaaand, it’s frickin expensive to just do basic scaling benchmarks. So I’ve started scaling some popular OSS models used in computer vision, which isn’t really getting much attention these days.

I’m starting with public vision data in microscopy, and the popular DINO, JEPA and SAM family of models. You can play around with my results in this cute little webapp, or pick up Tensorboard event logs and checkpoints from my HuggingFace at https://huggingface.co/DnaRnaProteins.

Open the full-page playground if you want more room to play with the web app.