text-to-text

Definition

Casting every task — classification, translation, summarisation, QA — as mapping an input string to an output string. T5 made this framing central, and the payoff was uniformity: one model, one loss, no task-specific heads, so tasks can be mixed freely during training. It is the direct ancestor of how instruction-following models handle heterogeneous work through a single generative interface, at the cost that structured outputs must be serialised and parsed back.