Ask HN: What are folks using to train/fine-tune Vision Language Models