Abstract

Large Language Models (LLMs) have seen a lot of interest from all over the world, especially since ChatGPT became the fastest growing consumer internet app in history. As we enter a new era of possibilities with AI, new challenges also present themselves. In July of 2023, Meta open-sourced the largest language models to date, making it one of the most important moments in the development of AI. Llama2 was the first LLM of its size and capabilities to be open-sourced; both the base LLM as well as a version fine-tuned for chat were released publicly for researchers to industry practitioners to leverage. In this talk, I will recap the journey of making Llama2 models safe and robust against misuse in hate speech, misinformation, etc. The talk will cover the technical details of how we defined what is safety for an LLM , the strategies we leveraged to train and fine-tune the models towards being safe, and the evaluations we conducted to verify that we had the level of safety we desired. I will also discuss the challenges that remain, and what the possible directions to address those are.

Link to join virtually: https://cam-ac-uk.zoom.us/j/81322468305

This talk is not being recorded.