An online system uses a model to detect violations of policies enforced by the online system for content uploaded to the online system by users for viewing by other users. The online system trains the model in multiple stages. To train the model, the online system obtains a set of training content items, with each content item of the set labeled with both a policy violated by the content item and a source of the content item, which acts as a proxy for a sub-category identifying a way in which the content item violated the policy. In the first stage, the online system trains the model using the set of training content items. In a second stage, the model of trained to predict policy violations from content items that are not labeled with a source. For example, the second stage is performed by freezing earlier layers in the model.