Most previous patent classification methods have treated the task as a general text classification task, and others have tried to implement XML (extreme multi-label learning) methods designed to handle vast numbers of classes. However, they focus only on the IPC subclass level, which has fewer than 700 labels and is far from "extreme." This paper presents a French Patents corpus INPI-CLS extracted from the INPI internal database. It contains all parts of patent texts (title, abstract, claims, description) published from 2002 to 2021, with IPC labels at all levels. We test different XML methods and other classification models at the subclass and group levels of the INPI-CLS dataset with about 600 and 7k labels, respectively, demonstrating the XML approach's validity to patent classification.