Andrew Ng and Sharon Zhou introduce a new course on fine-tuning large language models. Fine-tuning sits between prompting and pre-training from scratch, letting developers specialize models on domain data with just 100 to 10,000 examples in minutes to hours. The course covers instruction fine-tuning with hands-on code.
Hi, I'm happy to be back with Sharon Zhou, the instructor for this course, fine Tuning Large language models.
Hi Andrew, super excited to be back.
As you may know, Sharon had also previously taught DeepLearning AI's short course on diffusion models as well as the GAN specialization hosted on Coursera. In this course, Sharon will be diving into fine tuning OLMs, that is large language models. Broadly, there are three ways that application developers are using OLMs. First is prompting, which is really fast and can get you pretty good results. For example, you can write a prompt to tell an LLM to classify a piece of text as positive or negative sentiment and have that up and running in minutes. The other extreme is training, or sometimes we say pre training, a large foundation model from scratch that can cost tens of millions of dollars and require hundreds of billions or more words to train on. Most people outside well resourced companies might have a hard time doing that. Finally, there's an important design point in between which is to fine tune an open source language model.
You can take an LLM and fine tune it on only 100 examples, 1,000 or even 10,000 examples, and this takes minutes to hours rather than seconds with prompting or months with pre training. But this can also get you a significantly higher level of performance on your specific use case than either of the other approaches.
The world has quickly adapted to the powerful capabilities that LLMs like ChatGPT have made possible. But if you can't make use of your own domain specific or proprietary data, you can start to run into the limitations of the general purpose LLMs. With fine tuning you can specialize an LLM to your own data.
So in this course you'll learn all about 1 how fine tuning fits into training, 2 how it differs from prompt engineering or retrieval augmented generation alone and three you'll dive into a specific variant called instruction fine tuning that teaches an LLM to follow instructions similar to ChatGPT and you'll go through all of these steps in code. Then you'll come away from the course with a concrete understanding of fine tuning LLMs and be able to get started using these techniques for your own projects.
So lots of cool things in this short course and I think fine tuning is a significant step up in capability compared to just pre prompting and it is an important tool for people building applications using LLMs to know. So I hope you enjoy the course.
