Courses / Intermediate

Build your own AI

Choose data, train a model, control it, test it honestly and save it as a project you can share. 15 lessons in 4 modules, about 2 hours in all.

0 of 15 lessons done

Next up: Choosing your data

Your data

What to train on, how to clean it and how a tokeniser sees it.

3 lessons, 23 minutes

  1. 1Choosing your dataA model can only become what its data is. Decide what you want it to do, then pick text that shows exactly that, and enough of it.7 min
  2. 2Cleaning your textReal text is messy. Strip out the noise, make it consistent, and remove duplicates, in a few lines of Python you can run here.8 min
  3. 3How your model readsBefore training, text is split into tokens. See how the choice between letters and word pieces changes what a model can learn.8 min
  4. Module 1 quiz: Your data10 questions with diagrams to drag, slide and click. Pass with 12 to open what comes next.20 marks

Your model

Pick a size, run training and read the curves.

4 lessons, 32 minutes

  1. 1Picking a sizeMore parameters can learn more, and also memorise faster. Count them, and match the size of the model to the size of your data.7 min
  2. 2A first training runTrain a real transformer in this page and learn what to watch: the loss, the samples, and the moment it starts to memorise.9 min
  3. 3Reading the curvesA loss curve tells you whether to wait, stop or change something. Learn the five shapes you will see most, and what to do about each.8 min
  4. 4When it memorisesMemorising is the most common way a small model goes wrong. See it happen on a simple network, and learn four ways to prevent it.8 min
  5. Module 2 quiz: Your model10 questions with diagrams to drag, slide and click. Pass with 12 to open what comes next.20 marks

Controlling it

Temperature, prompts, saving, loading and fine-tuning.

4 lessons, 31 minutes

  1. 1Temperature and samplingThe same trained model can be careful or adventurous. One number, the temperature, controls which, and you can set it.7 min
  2. 2Steering with a promptThe beginning you give your model decides where it goes next. Use the Predict tab to steer it one character at a time.8 min
  3. 3Saving and loading a modelA trained model is just its numbers. Save them to a file, load them back, and carry on, in code you can run here.8 min
  4. 4Fine-tuning on new textInstead of starting from nothing, take a trained model and continue training it on a small, chosen set of text. It is faster and needs far less data.8 min
  5. Module 3 quiz: Controlling it10 questions with diagrams to drag, slide and click. Pass with 12 to open what comes next.20 marks
  6. Checkpoint 1: modules 1 to 325 questions with diagrams to drag, slide and click. Pass with 30 to open what comes next.50 marks

Shipping it

A second kind of model, honest testing, and a project of your own.

4 lessons, 32 minutes

  1. 1A second kind of modelNot every model writes text. Build a classifier that recognises your own drawings, the other half of what AI does every day.8 min
  2. 2Testing honestlyA model that looks good on its own training data proves nothing. Test on data it has never seen, and look at the kinds of mistakes it makes.8 min
  3. 3Your first projectPut it all together in a project that saves to your account: pick the data, train, test, write notes, and share it with a link.9 min
  4. 4What to build nextFrom a model with thousands of parameters to the ones with billions, and practical next steps for taking your own building further.7 min
  5. Module 4 quiz: Shipping it10 questions with diagrams to drag, slide and click. Pass with 12 to open what comes next.20 marks
  6. Case study50 questions with diagrams to drag, slide and click. Pass with 30 to open what comes next.50 marks
  7. Final paper50 questions with diagrams to drag, slide and click. Pass with 60 to open what comes next.100 marks