Building Data-Efficient and Reliable Applications with Large Language Models

Explore the challenges and solutions in developing reliable applications using Large Language Models (LLMs) in this 45-minute talk by Xinyi Cindy Wang from the Center for Language & Speech Processing at JHU. Delve into the complexities of adapting LLMs for specialized domains and personalized user needs, particularly in multilingual contexts. Discover data-efficient tuning methods and a novel factuality evaluation framework designed to address the limitations of LLMs trained on web data. Learn about Wang's research on multilingual instruction-tuning for Gemini and generative models used in Google search, as well as her contributions to data selection, representation, and model adaptation for multilingual natural language processing during her PhD at Carnegie Mellon University.