Wirewalk Instituteof AI
mod-310 Expert AvailableFlagship

Fine-Tuning LLMs: Full, LoRA and QLoRA

For ML engineers who need to decide whether to fine-tune an open-weight LLM, then design the data, size the GPUs, run full, LoRA or QLoRA training, and ship an evaluated adapter.

Level
Expert
Study time
20 hours
Lessons
18 in 6 modules
Exam
25 questions

About this course

Fine-tuning is the most over-requested and under-planned technique in applied AI. This course starts with the decision itself: what fine-tuning can and cannot change, and how to choose a base model and license you can actually ship. It then treats data as the main engineering artifact, covering chat templates, loss masking, deduplication and contamination.

The core of the course is mechanics with numbers attached. You will compute the memory budget for full fine-tuning byte by byte, derive LoRA's parameter count for a real model shape, and see exactly where QLoRA's NF4 quantization, double quantization and paged optimizers save memory. You will run jobs with TRL and PEFT, compare Axolotl and Unsloth, watch for overfitting and forgetting, build an evaluation that would catch a regression, and serve multiple adapters from one base model. The course closes with an overview of preference tuning with DPO. Product facts are current as of October 2026.

What you will be able to do

  • Decide between prompting, retrieval and fine-tuning for a given problem and justify the choice with the failure mode each one addresses
  • Select a base model by size, variant and license, including the obligations that open-weight licenses attach to fine-tuned derivatives
  • Design a supervised fine-tuning dataset with correct chat templates, loss masking, deduplication and contamination controls
  • Estimate GPU memory for full fine-tuning, LoRA and QLoRA from model shape, precision and optimizer, and choose hardware accordingly
  • Configure LoRA and QLoRA hyperparameters (rank, alpha, target modules, learning rate) and explain the effect of each
  • Run and monitor fine-tuning jobs with TRL/PEFT or a config-driven framework, recognizing overfitting, forgetting and divergence
  • Evaluate a fine-tuned model against its base and ship it by merging or serving adapters with vLLM
  • Explain how DPO uses preference pairs and a reference model, and when to add it after SFT

Syllabus

6 modules · 18 lessons · about 20 hours

Module 1: Deciding to fine-tune

What fine-tuning changes, which base model and license to start from, and whether to use a vendor API or your own GPUs.
3 lessons
  1. Fine-tune, prompt or RAG: a decision framework
  2. Choosing a base model: size, variant and license
  3. Vendor fine-tuning APIs versus self-hosted training

Module 2: Data is the model

Formatting, masking, cleaning and protecting the dataset that determines what the model learns.
3 lessons
  1. Chat templates, loss masking and dataset formats
  2. Data quality, deduplication and dataset size
  3. Contamination, leakage and a held-out eval set

Module 3: SFT mechanics and full fine-tuning

What supervised fine-tuning optimizes and the byte-level memory budget of updating every weight.
3 lessons
  1. What SFT optimizes: loss, packing and learning rate
  2. GPU memory math for full fine-tuning
  3. Making full fine-tuning fit: checkpointing, 8-bit optimizers and sharding

Module 4: LoRA and QLoRA

Low-rank adaptation from first principles, its hyperparameters, and 4-bit quantized training.
3 lessons
  1. LoRA math: low-rank updates, rank, alpha and scaling
  2. LoRA hyperparameters and target modules in practice
  3. QLoRA: NF4, double quantization and paged optimizers

Module 5: Tooling and running the job

Running jobs with TRL and PEFT, choosing a framework, and reading the signals of a healthy or failing run.
3 lessons
  1. TRL and PEFT end to end
  2. Axolotl, Unsloth and torchtune: choosing a framework
  3. Monitoring a run: overfitting, forgetting and divergence

Module 6: Evaluate, ship and go further

Proving the model improved, putting adapters into production, and the next step of preference tuning.
3 lessons
  1. Evaluating a fine-tuned model
  2. Merging, exporting and serving adapters
  3. Preference tuning with DPO: an overview

Exam and certificate

The exam unlocks once you have completed every lesson. It draws 25 questions from a larger bank, spread across the modules, with the answer options shuffled for each attempt. You have 60 minutes and need 80% to pass. After a failed attempt you can try again after 24 hours, up to 3 attempts in total. Feedback is given per module, never as the answer key.

Passing earns a certificate of completion and assessment from the Wirewalk Institute of AI with a unique ID that anyone can check at institute.wirewalk.com/verify. It is not an accredited qualification or a vendor certification. How certification works.

The Wirewalk Institute of AI is independent. This course teaches publicly documented products and is not affiliated with or endorsed by Hugging Face (Transformers, TRL, PEFT), bitsandbytes, Axolotl, Unsloth, PyTorch (torchtune), vLLM, OpenAI, Google Cloud Vertex AI, Amazon Bedrock, Meta Llama, Qwen, or Gemma. Product and company names are trademarks of their respective owners.

More in Model Development

All 16 courses in this track
mod-201 Intermediate Available

Deep Learning with PyTorch

A hands-on course in modern PyTorch: tensors and autograd, nn.Module design, data loading, a training loop you understand line by line, GPU and mixed-precision training, torch.compile, and exporting a model for deployment.

14 h · 19 lessons See price at checkout
mod-202 Intermediate Available

Transformers from the Ground Up

Build a decoder-only transformer piece by piece: tokenization, embeddings, scaled dot-product attention, multi-head and grouped-query attention, positional schemes such as RoPE, and the training and sampling loop, then map it onto real open-weight model architectures.

16 h · 19 lessons See price at checkout
mod-203 Intermediate Waitlist

Classical Machine Learning in Production

Gradient-boosted trees and linear models still win most tabular problems. This course covers leakage-proof pipelines, honest validation, tuning, calibration and explainability, then packaging, serving and monitoring for drift.

14 h · 19 lessons Join the waitlist