Pre-training with general-reason and activity-specific facts improves task functionality devoid of hurting other model abilities