language

Why we included this project

For anyone doing natural language processing research, this repository collects the runnable code behind dozens of papers from Google's language research group in one place. Alongside well-known work on retrieval-augmented question answering and the Canine tokenizer, it holds a broad mix of smaller studies covering compositional generalization, multilingual text, and probing how language models use the information they were trained on. Each subproject is self-contained, with its own data pipeline and training scripts, which makes it practical for reproducing a specific paper's results or borrowing pieces of an approach rather than adopting one integrated system. Because the collection spans several years of published research, it also works well as a study resource for tracing how these techniques evolved. Treat individual subdirectories as starting points and adapt them for your own experiments; none of this is production-ready software.