Explore options for fine-tuning Llama 3 models and dive into TorchTune, a library built to simplify fine-tuning. This chapter guides you through data preparation, TorchTune's recipe-based system, and efficient task configuration, providing the key steps to launch your first fine-tuning task.

The Llama fine-tuning libraries

Listing TorchTune recipes

Running a TorchTune task

Preprocessing data for fine-tuning

Filtering datasets for evaluation

Creating training samples

Saving preprocessed datasets

Fine-tuning with TorchTune

Defining custom recipes

Saving custom recipes

Running custom fine-tuning

Preparing for Llama fine-tuning

Learn how fine-tuning can significantly improve the performance of smaller models for specific tasks. Start with fine-tuning smaller Llama models to enhance their task-specific capabilities. Next, discover parameter-efficient fine-tuning techniques such as LoRA, and explore quantization to load and use even larger models.

Model fine-tuning with Hugging Face

Setting up Llama training arguments

Fine-tuning Llama for customer service QA

Evaluate generated text using ROUGE

Efficient fine-tuning with LoRA

Using LoRA adapters

LoRA fine-tuning Llama for customer service

Making models smaller with quantization

Loading 8-bit models

Speeding up inference in quantized models

Congratulations!

Fine-tuning with SFTTrainer on Hugging Face

Fine-tune the Llama open-source LLM to optimize its performance for specialized applications. Learn how to use TorchTune's recipe-based system and Hugging Face’s SFTTrainer to run fine-tuning loops, preparing your models with custom datasets. Finally, experiment with parameter-efficient techniques like LoRA and apply quantization to decrease model memory usage and run your tasks more efficiently.

<h2>Fine-tuning the Llama model</h2> This course provides a comprehensive guide to preparing and working with Llama models. Through hands-on examples and practical exercises, you'll learn how to configure various Llama fine-tuning tasks. <h2>Prepare datasets for fine-tuning</h2> Start by exploring dataset preparation techniques, including loading, splitting, and saving datasets using the Hugging Face Datasets library, ensuring high-quality data for your Llama projects. <h2>Work with fine-tuning frameworks</h2> Explore fine-tuning workflows using cutting-edge libraries such TorchTune and Hugging Face’s SFTTrainer. You'll learn how to configure fine-tuning recipes, set up training arguments, and utilize efficient techniques like LoRA (Low-Rank Adaptation) and quantization using BitsAndBytes to optimize resource usage. By combining techniques learned throughout the course, you’ll be able to customize Llama models to suit your projects' needs in an efficient way.

Working with Llama 3

Fine-tune Llama for custom tasks using TorchTune, and learn techniques for efficient fine-tuning such as quantization.

Fine-tuning with Llama 3

Fine-Tuning with Llama 3

Llama Fundamentals

Speeding up inference in quantized models

Fine-Tuning with Llama 3

Instruções do exercício

Exercício interativo prático