07-29, 13:15–13:45 (US/Eastern), Filmhouse
Imagine what you could do if you could tune Large Language Models (LLMs) with contributions from across your community. There are many popular open LLMs, just look at Hugging Face. Up until now, if a developer wanted to contribute a change to an open LLM, the only practical option was to publish a fork of the model which results in many forks but no version of the model having all or many of the community’s changes. In this session we learn about a new, novel technique using the InstructLab open source project to enable incremental, community contributions to instruction-tune a LLM.
The LAB: Large-Scale Alignment for ChatBots paper introduces a new, novel technique to allow incremental contributions to instruction-tune a LLM using a taxonomy-guided synthetic data generation process and multi-phase tuning process. A new open source project, InstructLab, is now available enabling the use of this technique. In this session, we will learn about the basic principles of the LAB technique and about the new InstructLab open source project for applying the LAB technique. The InstructLab project allows the community to incrementally contribute new skills and knowledge in an open source manner which is then aggregated and used as the basis for the generation of synthetic training data that is then used to instruction-tune the LLM. This process can be repeated to continually expand the capabilities of the LLM over time.
I'm a Senior Technical Staff Member in IBM Research working on open source development and open technologies.