Explore a comprehensive technical lecture that delves into Meta and NYU's groundbreaking research on Self-Rewarding Language Models, focusing on eliminating the need for human-labeled data by enabling models to act as their own judges. Learn about the challenges of human-labeled data, the concept of super-human agents, and the intricate workings of self-rewarding language models through detailed explanations of instruction following and LLM-as-a-Judge capabilities. Discover the technical aspects of model initialization, dataset creation, self-instruction processes, and AI Feedback Training (AIFT) methodologies. Examine the evaluation methods and results that demonstrate the effectiveness of this innovative approach to language model training. Perfect for AI researchers, machine learning practitioners, and anyone interested in cutting-edge developments in natural language processing and artificial intelligence.
Deep Dive Into Self-Rewarding Language Models - Training Models as Their Own Judges
Overview
Syllabus
What we’re covering
The Problem With Human-Labeled Data
Super-human Agents and Synthetic Data
What is a Self-Rewarding Language Model
Skill 1. Instruction Following
Skill 2. LLM-as-a-Judge
Prompting as the Judge
Initialization and Datasets
Self-Instruction Creation
AI Feedback Training Data Creation AIFT
Iterative Training
Evaluation
Results
Conclusion
Join us!
Taught by
Oxen
Related Courses
-
Fine-Tuning Self-Rewarding Language Models with Mistral 7B
-
Improving Accuracy of LLM Applications
-
Creating Self-Instruct Data Sets for LLM Fine-Tuning with ChatGPT
-
Teaching Language Models to Use Tools - Deep Dive into Toolformer
-
Understanding How Llama 3.1 Works - A Technical Deep Dive
-
Self-Instruct Fine-Tuning of Large Language Models - Introduction to Alpaca
