Fine-Tuning LLMs (Lite)
Learn when fine-tuning beats RAG and prompting, why LoRA is the default, and how to run, evaluate, and check a small LoRA fine-tune for forgetting.
What You'll Learn
Understanding Why acme-assist Might Need Fine-Tuning
By now acme-assist answers from the help centre, with citations.
Deciding Whether You Need to Fine-Tune
Fine-tuning costs data, training time, and an artifact to maintain forever, so the default answer is no. This topic gives you the test to apply first.
Understanding What Fine-Tuning Changes
Fine-tuning moves a model's weights toward the examples you show it.
Understanding LoRA and QLoRA
LoRA is the technique you will actually use. A little intuition about it makes the three settings you tune easy to reason about.
Preparing a Small, Clean Dataset
The dataset is the real product of a fine-tune.
Training a LoRA Adapter
Training is a short, mostly mechanical step once the data is right. The goal here is to read the script until nothing in it surprises you.
Skills You'll Master
Curriculum Index12 topics
Understanding Why acme-assist Might Need Fine-Tuning
By now acme-assist answers from the help centre, with citations.
Deciding Whether You Need to Fine-Tune
Fine-tuning costs data, training time, and an artifact to maintain forever, so the default answer is no.
Understanding What Fine-Tuning Changes
Fine-tuning moves a model's weights toward the examples you show it.
Understanding LoRA and QLoRA
LoRA is the technique you will actually use.
Preparing a Small, Clean Dataset
The dataset is the real product of a fine-tune.
Training a LoRA Adapter
Training is a short, mostly mechanical step once the data is right.
Evaluating the Adapter Honestly
A fine-tune that you have not compared against a good prompt is a guess.
Shipping and Versioning an Adapter
An adapter that works today but cannot be reproduced next month is a liability.
Hands-On Lab: Train and Judge a Tone Adapter
This lab trains one small adapter and then decides, with evidence, whether it earns its place.
Quick Reference
A short summary of the decisions and settings you will reach for most. Decision table Settings and checks
Common Mistakes
These are the five errors that waste the most time in practice, each with its fix.
Reviewing What You Built and What Comes Next
You started with a model that had the right facts and the wrong voice, and you finished with a measured answer to...
Career Impact
Roles that use the skills in this module.
AI Engineer
MLOps Engineer
Platform Engineer
Next Modules
Practice on the Coding Sheet
Not a software engineer sheet. Every problem comes from real DevOps, SRE, Platform and Cloud interviews, from your first script to a system you build yourself.
Open the Coding SheetFrequently Asked Questions
Use RAG when the model is missing facts or the facts change often, because you can update a document without retraining. Fine-tune only when the facts are right but the tone, format, or style is wrong and prompting has already failed. Most problems that look like fine-tuning problems are really prompting or retrieval problems.
LoRA freezes the original model and trains a small add-on called an adapter. It needs far less memory than full fine-tuning, produces a file of megabytes instead of gigabytes, and leaves the base weights untouched. That makes it the sensible first choice for almost every team.
For a narrow style change, a few dozen clean, consistent examples are enough to see the effect, and a few hundred is a more realistic starting point for production. Quality matters more than volume, because every bad example teaches the wrong pattern. Always keep part of your data aside to test on.
Yes, for a very small model such as a 0.5 billion parameter one, but it is slow on a laptop CPU. A free Colab GPU finishes the same job in minutes, so use Colab as the default for the lab in this module.
It is the loss of general ability after training a model hard on a narrow dataset. LoRA lowers the risk because the base weights stay frozen, but it does not remove it. Test the tuned model on prompts unrelated to your task before you ship it.