ORPO: Monolithic Preference Optimization without Reference Model
Offered By: Yannic Kilcher via YouTube
Course Description
Overview
Explore a comprehensive analysis of the ORPO (Odds Ratio Preference Optimization) algorithm, a groundbreaking approach to language model preference alignment without the need for a reference model. Delve into the paper's key findings, which demonstrate how ORPO eliminates the necessity for a separate preference alignment phase in language model training. Examine the empirical and theoretical evidence supporting the effectiveness of the odds ratio in contrasting favored and disfavored generation styles during supervised fine-tuning. Learn how ORPO, when applied to models like Phi-2, Llama-2, and Mistral, achieves state-of-the-art performance on benchmarks such as AlpacaEval2.0, IFEval, and MT-Bench, surpassing larger language models. Gain insights into the crucial role of supervised fine-tuning in preference alignment and understand how ORPO's innovative approach simplifies the process while maintaining high performance.
Syllabus
ORPO: Monolithic Preference Optimization without Reference Model (Paper Explained)
Taught by
Yannic Kilcher
Related Courses
Introduction to Artificial IntelligenceStanford University via Udacity Natural Language Processing
Columbia University via Coursera Probabilistic Graphical Models 1: Representation
Stanford University via Coursera Computer Vision: The Fundamentals
University of California, Berkeley via Coursera Learning from Data (Introductory Machine Learning course)
California Institute of Technology via Independent