PyTorch Expert Exchange: Adapting open source models with Open-Instruct and Tulu
Abstract: While language models are powerful tools, they require adaptation (often called ‘post-training’) to dramatically increase their usability and performance. Hamish Ivison - PhD student at the University of Washington at H2Lab will talk about Ai2’s efforts in post-training by tracing the evolution of the open-instruct library, from a project-centric academic repository to a (more) fully-fledged post-training codebase, and highlight how it serves our post-training team. In addition, Hamish will discuss how open-instruct has powered the Tulu project, a series of state-of-the-art models adapted from llama models, including a sneak peek at the upcoming Tulu 3 release.
