Atto is a family of compact contextual models designed for parameter-efficient language representation learning.