Fine-Tuning with Llama 3
| Course Details | ||||||||
|---|---|---|---|---|---|---|---|---|
|
||||||||
| Languages | ||||||||
|
||||||||
| Outcomes | ||||||||
|
||||||||
| Price & Terms | ||||||||
|
||||||||
| Technical | ||||||||
|
||||||||
| Specification | ||||||||
|
||||||||
Prepare data, run TorchTune recipes and fine-tune Llama 3 models with SFTTrainer, LoRA and 8-bit quantization to save memory.
Fine-tuning adapts a Llama 3 model to one task, such as answering customer service questions, and it can markedly lift what smaller models achieve on that task. You turn a dataset into training samples, launch fine-tuning runs with ready-made or custom recipes, and cut memory needs with low-rank adapters and 8-bit loading so that larger models fit on limited hardware.
What You Will Achieve
- Turn a raw dataset into filtered, saved training samples ready for a Llama fine-tuning run
- Configure and launch a fine-tuning task with a built-in or custom TorchTune recipe
- Set training arguments, fine-tune a smaller Llama model with SFTTrainer and score its output with ROUGE
- Reduce memory use by training LoRA adapters and loading models in 8-bit with quantization
Program Overview
Data Preparation and TorchTune Recipes
Surveys the libraries available for Llama fine-tuning, then works through TorchTune, its recipe system and the data steps that come before a run.
- Overview of Llama fine-tuning libraries
- Listing and running TorchTune recipes
- Preprocessing, filtering and saving datasets
- Defining and saving custom recipes
Fine-Tuning with SFTTrainer, LoRA, and Quantization
Moves to Hugging Face, where a small Llama model is trained for customer service question answering and then made lighter with parameter-efficient methods.
- Training arguments for SFTTrainer
- Customer service QA fine-tuning and ROUGE evaluation
- LoRA adapters for efficient training
- 8-bit loading and faster quantized inference
Certificate Awarded (Paid)
Job Outlook
Fine-tuning skills apply wherever teams adapt language models to narrow jobs, such as customer service chatbots, domain-specific text generation, summarization, translation and question answering.
Good for: Practitioners who want to adapt a Llama 3 model to a single task, such as customer service answers, while keeping memory use down during training.
Keep in mind: The two chapters center on TorchTune and Hugging Face’s SFTTrainer, and the one evaluation method taught is ROUGE scoring of generated text.
Bottom line: An exercise-led route through two fine-tuning toolkits that ends with LoRA and quantization, the methods that let larger Llama models load and run in less memory.
Frequently Asked Questions
How do TorchTune recipes work in the lessons?
TorchTune is a library built to simplify fine-tuning Llama models. The first chapter uses its recipe-based system: you list the built-in recipes, run a task, then write and save custom recipes to launch your own fine-tuning run.
How is the fine-tuned model evaluated?
After a Llama model is fine-tuned for customer service question answering, the lessons score its generated text with ROUGE. The first chapter also covers filtering a dataset to set data aside for evaluation.
What does quantization change about a Llama model?
Quantization makes a model smaller so it needs less memory. The lessons load models in 8-bit with BitsAndBytes, which lets larger Llama models be used, and show how to speed up inference once a model is quantized.
User Reviews
Write your review
You must be logged in to post a review.
Reviews can only be posted by registered learners. Log in or create a free account, then you will return here to post.
There are no reviews yet.