I always recommend kaggle for its great variety of challenges and applicability to real world scenarios (specially the competitions).
Apart from that, a good application you can enjoy developing is a real-time object detector, like yolov5. You can even collect your own data, label it and fine tune the model with your specific training.
Try the UCI Machine Learning Repository. They have hundreds of free data sets. Download a set and try to build a classifier from it. If it is a multi-class set, you can build the multi-class (softmax) classifier and also several binary classifiers.
If you have a data set of animal images, for example, you could build dog vs cat, dog vs horse, cat vs horse, dog vs not-dog, etc.