This paper introduced a method to train models (like GPT-3) to summarize text by using Reinforcement Learning from Human Feedback (RLHF) . 📂 What is in the ZIP?
The identifier qsUa0c4PEVK2XcJiGiow is specifically used by and GitHub for the official release of their human preference data. It typically contains: Thousands of comparisons between model-generated summaries. Rankings provided by human labelers. Data used to train the "Reward Model" that powers RLHF.
Stiennon, Ouyang, Wu, Ziegler, Lowe, Voss, Radford, Amodei, and Christiano. Organization: OpenAI.
Neural Information Processing Systems ( NeurIPS 2020 ).