Model CI/CD Course: LLM case study overview

Key Takeaways

hello welcome back so far in this course you've learned about model registry specific types of automation including web Hooks and weights and bu's launch and now we want to help you apply these tools to a real project and for this my colleague darck is going to be walking you through a case study and dar is actually a really impressive person so first of all uh Dar is a kaggle Grandmaster um he he is really good at machine learning but also he has uh many years of experience in Industry applying machine learning and production and also he's one of my favorite teachers he's uh taught courses with me in the past uh especially courses on weights and biases and so I really think you're going to enjoy this lesson thanks a lot ham I I really appreciate this introduction so again I'm a machine learning engineer at weights and biases one thing that we do at weights and biases is we work a lot with our customers a lot of customers are now very excited about training and fine-tuning llms uh in fact we have a separate course about training and fine-tuning llms uh that we will link uh below this video and that we definitely recommend everyone taking but we also want to take this this use case of fine-tuning um a large language model into this course and show how to set up U Model Management processes that that uh will support uh both training and evaluation of uh of uh llm models and um since this is a case study we will take an existing data set called alpaka it comes uh from uh from Stanford it's a it's a synthetic data set that was created with um N llm Text D Vinci 003 and it contains um uh several thousand of like in fact 5050 around 50,000 of instruction following examples and in the original alpaka paper uh the team has Fine tuned a 7 billion llama model uh on this data set in our case since we are trying to be more efficient here and we want to also make it easier for people following the course will be fine tuning a smaller a tiny Lama model that has around 1 billion parameters the important thing when when fine-tuning llms is evaluation and in this case study we'll um we'll uh build on the concept of using llm as a judge so uh we will take um the generations from the model that we fine tune we'll compare them to Generations from a model that uh we we defined as our Baseline and we will use uh another llm in this case a more powerful llm which is gp4 to compare which one is better and this will result in a metric that we will use to make decisions if uh the new model that we trained is better than the Baseline and if you want to proceed further with uh quantizing it preparing it for deployment and ultimately deploying it into production and with this I want to now transition into the the live uh the live demo session and I want to encourage you to work along with us uh check out the repo that we are uh we will provide for the exercises follow along and practice uh setting up this Model Management processes uh together with us