Case study: Faster fine-tuning with dataset clustering and thinning