Fine-Tuning LLMs with LoRA, QLoRA and PEFT

Course Overview

Advanced

Free Course

For engineers who can already call and prompt a language model and now need to decide whether, and how, to adapt an open model to their own task. You will learn to work out the memory and cost of a fine-tune before you start, train a LoRA or QLoRA adapter on a single GPU, and prove with a proper evaluation that it beats the base model.

Instructor: Jaidev

Sections: 4

Course Content

Section 1: Understanding Fine-Tuning Paradigms

Work out what full fine-tuning and PEFT really cost, decide whether fine-tuning is the right fix at all, and build a dataset worth training on.

1. Full Fine-Tuning vs. Parameter-Efficient Fine-Tuning (PEFT)

2. When and Why to Fine-Tune — and When Not To

3. Dataset Design, Quality and Labelling Strategies

Section 2: LoRA and QLoRA Techniques

Learn how low-rank adapters and 4-bit quantisation work, compute their memory cost by hand, and run a complete QLoRA training job.

1. LoRA Architecture — Low-Rank Adapters

2. Quantisation — FP32, FP16, BF16, INT8 and INT4

3. Training with QLoRA — An End-to-End PEFT Pipeline

Section 3: Practical Fine-Tuning Pipeline

Turn a one-off notebook into a reproducible project, tune a run and keep its best checkpoint, and judge the result against the base model.

1. Project Structure and Configuration Management

2. Hyperparameters and Checkpoints — Tuning a Run

3. Evaluation and Alignment — Is the Model Good?

Section 4: Mini Project

Fine-tune a 7B model with QLoRA on one 16 GB GPU, end to end, and show honestly whether it beats the base model.

1. Mini Project: Fine-Tune a Domain-Specific LLM with QLoRA