Pre-training with both group and individual human context significantly improves user-level regression tasks and document-level classification tasks.