In this article, we learn about the Chain of Density method for text summarization and how to implement it using Instructor. The Chain of Density method involves iteratively refining a summary by adding missing article entities. Salesforce AI Research has found that this method consistently outperforms human-generated summaries. By using Instructor, we can distill a GPT 3.5 model to match the capabilities of GPT-4’s iterative summarization. This approach reduces latency by 20x, lowers costs by 50x, and maintains entity density. The implementation details, including code examples and validators, are provided for reference.
https://jxnl.github.io/instructor/blog/2023/11/05/chain-of-density/