For a long time, facial feature point detection has been plagued by problems such as occlusion and posture changes. The detection task is no longer seen as a single independent problem, and researchers try to improve the robustness of detection through multi-task learning. Therefore, the MAFL dataset came into being. MAFL stands for Multi-Attribute Facial Landmark, which consists of manually annotated feature points of 19,000 training images and 1,000 test images, and is used for facial feature point detection research based on deep multi-task learning.