Large-Scale Pretraining: The Nitty-Gritty Details

This talk will give a rare close-up of the nitty-gritty details that go into training large-scale LLMs. In the autumn of 2023, Aleph Alpha Research Lab prepared to train their next generation of large language models, which are training now.

In this talk, Robert Baldock will chronicle their learnings from this process. In particular, he will describe their experiments to optimise the architecture and pretraining, their optimal scaling study, insights about efficient and numerically stable parallel training, tokenizer construction, and the preparation of the large-scale web-crawl dataset.