T5
Google's 2019 encoder-decoder model that framed every language task as turning one text into another.
Also: Text-to-Text Transfer Transformer
T5, published by Colin Raffel and colleagues at Google, treated translation, summarisation, classification and question answering alike as text in, text out, with a short prefix naming the task. Its paper compared many training choices systematically, and the model was trained on C4, a cleaned 750-gigabyte slice of Common Crawl that became a widely used dataset. The largest T5 had 11 billion parameters. Later variants include the multilingual mT5, covering over 100 languages, and the instruction-tuned Flan-T5.