Selective Reflection-Tuning is a novel approach that enhances instruction-tuning data by leveraging a teacher model's reflection and introspection to improve data quality and compatibility with the student model. This method involves a teacher-student collaboration where the teacher model generates improved instruction and response pairs, and the student model selects the most beneficial ones based on metrics like Instruction-Following Difficulty (IFD) and reversed-IFD (r-IFD). The process ensures that the data is refined to better suit the student model's capabilities, leading to more efficient and effective instruction tuning. The method does not require new data collection, instead using existing data to enhance the quality through a selective recycling process. Experiments show that this approach significantly improves the performance of large language models, outperforming existing models with smaller datasets. The method is versatile and can be applied to various instruction-tuning scenarios, demonstrating its effectiveness in enhancing data quality and model performance. The key contributions include the introduction of a teacher-student collaboration pipeline, the development of a nuanced evaluation schema using IFD and r-IFD, and the demonstration of superior performance with minimal data. The approach is efficient, requiring only instruction tuning on a small amount of data, and is effective in improving the quality of instruction-tuning datasets.Selective Reflection-Tuning is a novel approach that enhances instruction-tuning data by leveraging a teacher model's reflection and introspection to improve data quality and compatibility with the student model. This method involves a teacher-student collaboration where the teacher model generates improved instruction and response pairs, and the student model selects the most beneficial ones based on metrics like Instruction-Following Difficulty (IFD) and reversed-IFD (r-IFD). The process ensures that the data is refined to better suit the student model's capabilities, leading to more efficient and effective instruction tuning. The method does not require new data collection, instead using existing data to enhance the quality through a selective recycling process. Experiments show that this approach significantly improves the performance of large language models, outperforming existing models with smaller datasets. The method is versatile and can be applied to various instruction-tuning scenarios, demonstrating its effectiveness in enhancing data quality and model performance. The key contributions include the introduction of a teacher-student collaboration pipeline, the development of a nuanced evaluation schema using IFD and r-IFD, and the demonstration of superior performance with minimal data. The approach is efficient, requiring only instruction tuning on a small amount of data, and is effective in improving the quality of instruction-tuning datasets.