Skip to content
Local LLM Course

Machine Learning Foundations for LLM Work

Readers arrive at this course with a command line and curiosity, not a statistics degree. That is fine. Everything from “did my fine-tune help” in Part 13 to “why is the 4-bit model slower on this box” in Part 6 depends on five ideas, and they are cheaper to teach once, properly, than to re-explain in twenty later lessons.

This part is those five ideas. What learning is, mechanically: a loss, a gradient, a step. How a network is built from weighted sums and why depth needs residual connections. How to tell a model that learned from one that memorised, which is the method every evaluation in this course uses. What a GPU actually accelerates and what each number format gives up. And then a lab in which you set up the toolchain on your own machine and train a small network end to end, watching the loss fall and the validation curve turn.

What you will be able to do

Section titled “What you will be able to do”

By the end of this part you should be able to:

  • describe a training run as a loop, name each stage, and say which hyperparameters control it;
  • explain what a neuron, a layer and a deep network are, what backpropagation computes and why training needs so much more memory than inference;
  • design a train, validation and test split that can detect memorisation, read a pair of loss curves, and name the ways an evaluation can leak;
  • say what a tensor is, why matrix multiplication dominates, why memory bandwidth rather than compute limits most of what this course does, and what FP32, BF16, FP8 and 4-bit formats each give up;
  • set up Python, a virtual environment and the right PyTorch or MLX build for your platform track, verify the accelerator is in use, and train and evaluate a small model.

What you need first

Section titled “What you need first”

Nothing beyond the prerequisites on the home page: comfort with a command line and some Python. No model larger than a few hundred megabytes is downloaded in this part, and every page runs on the 8 GB tier. If your machine has no accelerator at all, the lab still runs on the CPU; it is just slower, and the lab says by how much.

How to work through it

Section titled “How to work through it”

Read the four lessons in order; each uses the previous one’s vocabulary. Then do the lab with a terminal open. The lab is the first place the four platform tracks appear, and it asks you to start the lab notebook you will keep for the whole course. Do not skip that: the capstone is built from it.

Part1 of 27Level1 — AI LiteratePages5Estimated time2.8 hours

0 / 5 lessons in this part completed