PAPER 27 · Bonus · text-to-text

T5: Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer

Raffel et al. 2019 Paper

Reframed many NLP tasks as text-to-text generation under one unified training format.

Core concept

T5 turns many different NLP tasks into one format: text goes in, text comes out.

Why it mattered

It made task unification a major training and product design idea.

Visual shortcut · Everything as text-to-text
task prompt
text input
text output
one format

T5 makes many NLP tasks share one interface: tell the model the task, give text, get text back.

How it works
Convert each task into text input.
Add a prefix that names the task.
Train one model across tasks.
Generate every answer as text.

The quick digest

Before T5, NLP tasks often had custom formats: classification labels here, spans there, translations somewhere else. T5 says: make everything text-to-text. If the task is sentiment, output text. If it is translation, output text. If it is summarization, output text.

That unification simplifies transfer learning. The same model and training interface can cover many tasks, with task prefixes telling the model what kind of transformation to perform.

The modern echo is obvious: prompts are text instructions that turn different workflows into a shared interface. T5 is one of the bridges from classic NLP task engineering to instruction-style model interaction.

What to remember

One-liner
Everything becomes text in, text out.
Why it matters
A unified interface simplifies transfer.
Builder instinct
Task prefixes foreshadow modern prompting.

Read it like this

Build instinct

Turn three different tasks into text-to-text examples and train or prompt one model through the same interface.

Read source → All papers
Previous26 · The Smol Training Playbook