Building Instruction-Tuning Datasets from Human-Written Instructions with Open-Weight Large Language Models