Flan-T5

Published October 2022 in 'Scaling Instruction-Finetuned Language Models' (arXiv:2210.11416), Flan-T5 took Google's T5 encoder-decoder architecture and fine-tuned it on 1,836 tasks phrased as natural-language instructions — including chain-of-thought data. The result: models from 80M to 11B parameters that could follow unseen instructions zero-shot, with the method scaling to the 540B Flan-PaLM.

Flan-T5 established instruction tuning as the standard post-training recipe (the direct ancestor of RLHF-era alignment), and the checkpoints remain workhorses for research and efficient deployment — small enough to run anywhere, Apache 2.0 licensed, and still competitive on structured NLP tasks. Available in Small, Base, Large, XL, and XXL sizes via Hugging Face.