Self-driving technology using deep learning has achieved a lot of research and excellent performance, but the process of training a model requires a large amount of high-quality data and it is still difficult for humans to directly inspect the processed data and prove its quality. In addition, classifying the training difficulty of the data for contextual and functional separation of deep learning algorithm is also an important challenge. This paper proposes a framework for validating the quality of constructed new datasets, which also classifies the training difficulty of datasets, thereby ensuring the versatility of valid datasets and introducing strategies to classify contextual datasets. Experiments on the AI hub dataset proved its quality and were able to be reorganized into datasets classified by difficulty level.