This is the feature alignment pre-training work to train only only the multi-modal projector. "Predict" paragraph given caption, ocr and image token