This repository is Google’s official codebase for T5, the Text-to-Text Transfer Transformer that reframed every NLP task — translation, summarization, classification, question answering — as text going in and text coming out. It has earned about 6,600 stars on GitHub, is Apache 2.0 licensed, and hosts the original models and training code behind one of the most influential papers in modern NLP.
Before T5, each NLP task tended to have its own architecture and output format. The T5 paper, “Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer,” showed that a single encoder-decoder transformer pre-trained on the large C4 corpus could be fine-tuned for state-of-the-art results across benchmarks simply by formatting tasks as text-to-text problems with task prefixes. This repository contains the original Mesh TensorFlow implementation, pre-trained checkpoints in sizes from roughly 60 million to 11 billion parameters, and the fine-tuning recipes from the paper; it is also home to mT5, the multilingual variant covering 101 languages. While most practitioners today use T5 through Hugging Face, this repo remains the canonical source for reproducing the original research and understanding the foundations that later encoder-decoder and instruction-tuned models built on. Its design choices — relative position biases, the span-corruption pre-training objective — still show up in modern architectures.
Free. The T5 codebase and checkpoints are open source under the Apache 2.0 license; there is no paid version.
NLP researchers and engineers who want the canonical T5 implementation — to reproduce results, study large-scale transfer learning, or fine-tune encoder-decoder models from the original checkpoints.