Module 1Decide whether to adapt
2 lessonsDecide whether to adapt at all. Find out what exactly fails, compare a better prompt, retrieval, tools and fine-tuning against that need, and set targets and limits before anyone trains anything.
Fine-Tuning and Adapting AI Models
Decide with evidence whether to fine-tune a model, and if you do, do it so that you can trust and maintain the result. You prepare clean training data, train a small LoRA adapter on an ordinary computer, compare it with the prompt-only system on the same held-out tickets, and decide whether it ships.
- 12 lessons
- 1 case studies
- 10–14 hr
- Advanced
- Check in every lesson
- Free certificate
Read every lesson as a guest, or sign in to enroll and track your progress.
Enroll to save the lessons that you complete. It is free.
What you'll learn
By the end of this course, you will be able to:
- Choose adaptation only when justified by evidence.
- Prepare permissioned, high-quality training data.
- Run or inspect a small adapter-based experiment.
- Evaluate, version and deploy an adapted model responsibly.
About this course
Larkfield’s support assistant works. It uses a large hosted model with a long prompt, and the evaluation course measured it carefully. Then Tomás has an idea: “Let’s fine-tune our own model. It will be cheaper, faster and private, and it will know our policies.” It sounds good. But is it true, and how would anyone know?
This course is that decision, argued with evidence one step at a time. Every lesson is short. It teaches one idea, and then the team discusses a real choice: Tomás wants to fine-tune, Priya asks for evidence, Kwame asks what it costs and Grace asks what customers will see. You answer first, and then you read the course’s answer, argued from measured results.
You start with the failure: what exactly goes wrong, and would a better prompt, retrieval or a tool fix it? Then you prepare training data with a clear origin and permission, and you keep the test tickets out of it. You see how a LoRA adapter trains a small part of a model, and you train one on an ordinary computer. You compare it with the base model, a better prompt, the hosted model and a simple classifier on the same held-out tickets.
The answer can be “do not adapt”. That is a valid result when the evidence says so, and the course shows you how to write it down. If the adapter is worth keeping, you package it with everything it depends on, so that someone else can load it, check it and roll it back.
Before you start
You completed Deep Learning and Transformers Explained, Building Reliable Applications with LLM APIs and Evaluating and Testing AI Systems, or you have the same skills. You do not need a GPU, an account or money. A recorded training run, its adapter and every prediction are supplied, so every lesson works without training; training again is optional and runs on a laptop CPU.
Courses to take first: Deep Learning and Transformers Explained, Building Reliable Applications with LLM APIs, Evaluating and Testing AI Systems.
Who this is for
People who have trained a small neural network, built an application on a model API and evaluated it on a held-out set, as in Deep Learning and Transformers Explained, Building Reliable Applications with LLM APIs and Evaluating and Testing AI Systems, and who now hear "we should fine-tune our own model". Data scientists, AI engineers and developers.
How we made this course
Experienced AI practitioners designed this course. AI helped us draft the content; people chose the teaching approach and the examples, and reviewed every lesson. The outputs that you see are the actual results of running the code. How we make our courses
Inside the course
Module 2Prepare trustworthy training data
2 lessonsPrepare training data you can trust. Give every example one format, record where it comes from and what you may use it for, remove duplicates and wrong labels, and keep the test cases out of the training data.
Module 3Understand adaptation methods
2 lessonsUnderstand what changes when you adapt a model. Compare full tuning with a small LoRA adapter, count what each one trains in a real model, and estimate the memory, with quantization and other methods in their place.
Module 4Run a controlled experiment
2 lessonsRun one small, controlled experiment. Pin every version, plan the time and memory, train an adapter on the CPU, read the training and validation loss, and resume a stopped run from a checkpoint.
Module 5Evaluate beyond training loss
2 lessonsJudge the adapter by more than its training loss. Compare it with the base model, a better prompt, the hosted model and a simple classifier on the same held-out tickets, look for regressions, and weigh time, memory and cost.
Module 6Package and maintain the adaptation
2 lessonsPackage the result so that someone else can use, check and replace it, or write down why it does not ship. Version the adapter with everything it depends on, load and verify it, and plan how it is maintained.
Final module · choose at least oneCase studies
1 case studiesOne end-to-end case study on open data. Decide whether a team should fine-tune, with a prompt-only baseline, a small adapter and a written decision memo. Complete it for the certificate; it is a portfolio piece.
Complete at least one case study for the certificate. The others are optional. What each one covers
After the case studiesFinal assignment
Use every module in one practice project. You collect evidence, and check it against the course outcomes.
- Final assignment180 min
Case studies
Realistic projects from start to finish, on real data, with every output explained. Each one makes a portfolio piece. Complete at least one for the certificate; the others are optional.
Should this team fine-tune?2 hr 30 min
End to end on a public dataset with a clear licence: a prompt-only baseline, a small adapter trained on the CPU, the same held-out comparison and a written decision memo.
Reviews
Loading the reviews…
Took this course? Sign in to review it. Learners can write a review after they complete 25% of the lessons.