For these small models, you would fine tune them on your use case. https://huggingface.co/transformers/v4.8.2/training.html