Explore model merging, task-vector transport, and configurable fine-tuning for vision and text models with fast checkpoint evaluation