Let's build a real-world ML app with LLMs

Do you want to learn how to build real-world ML apps, using Large Language Models Then this FREE tutorial is for you.

The Hands-on LLMs Tutorial

In the Hands-on LLM tutorial you will learn how to build, step-by-step, a stock market advisor. You will do that by combining MLOps best practices, with the latest advancements in Large Language Models.

This is not about building a demo, but a fully working ML system following the 3-pipeline architecture.

The 3-pipeline architecture

MLOps becomes way simpler when you realize every ML system is made of 3 pipelines. This design scales easily from 1 developer (ML ninja) to a full MLOps team, allowing easy collaboration between your data engineers, data scientists, and ML engineers.

These are the 3 pipelines we will build in this course.

1. Feature Pipeline

The Feature pipeline prepares the data and makes it available to your models. In this tutorial you will build a feature pipeline that

• fetches financial news in real-time from the Alpaca News API,

• generates text embeddings using Bytewax and a Base Language Model

• stores the embeddings in a Serverless Verctor DB, in this case Qdrant.

2. Training pipeline

The Training pipeline fine-tunes a base LLM (in our case Falcon 7B) using a custom dataset we created for this project.

In this previous post I explained a bit how the dataset was generated.

We will log experiments and prompts using CometML, to keep track of what works best for this problem.

3. The Inference pipeline

The Inference pipeline makes your model predictions available to downstream apps or users.

In this case, we will deploy the fine-tuned model as a Serverless REST API with Beam, so predictions will be served on a request-response basis to clients, using Retrieval Augmented Generation (RAG) with the embeddings from the Vector DB.

Meet the instructors

Your instructors for this course are myself (what a surprise 😛) and Paul Iusztin, the lead developer of the project.

Paul has a great weekly newsletter on MLOps, Decoding ML, that I recommend you subscribe to.

Also, special thanks to Alexandru Răzvanț as well, who is helping on the development side of the course.

Let’s get started

In the following weeks we will release a series of video lectures that will guide you step-by-step through the whole development, from A to Z.

In the meantime, go check the repository yourself.

→ Give it a star ⭐ on GitHub

See you next week. Same place. Same time.

Peace and Love

Pau