Text this: Selective imitation for efficient online reinforcement learning with pre-collected data