A “backbone” model is designed to extract task-agnostic representations from specific data modalities (e.g., images, text, speech). This representation can then be fine-tuned for specialized tasks.