On: PaLI-3 Vision Language Models

[buildbot]

To finetune on each benchmark? I'd say it's not in our modern era of in-context learning, though of course fine-tuning has it's place as well for making smaller models better in one domain than a generalist larger model.