raiyanyahya/how-to-train-your-gpt

Builds a modern LLM from scratch, line by line, with every decision explained in plain language. Not a tutorial that waves hands at the hard parts — the repo covers tokenization, attention, training loops, and the bits that usually get glossed over in blog posts. What makes it worth your Saturday: the "explained like we are five" promise actually holds. If you have shipped software but never gone deep on how a transformer actually trains, this is the fastest path from confused to confident that does not require a $3,000 Coursera subscription. The annotation density is the real differentiator over the dozen other from-scratch repos — you are not reverse-engineering someone else's choices. Honest reservation: this is educational scaffolding, not production code. Do not pull it into a real project. Use it to stop nodding along in architecture meetings and start having opinions. -> Best for: AI engineer or technical PM who ships on top of LLMs but wants to understand what is actually happening underneath