Fine-Tuning AI Models

Over the past three years I have watched the same scene play out again and again: a team picks a good language model, spends two weeks wrestling with prompts and a knowledge base, and finally concludes that "the model doesn't understand our work." Then they turn to fine-tuning — and that is where they get lost.

Getting lost is not without reason. Persian-language resources in this field are either shallow, or hasty translations of tutorials that had themselves gone stale. Good English resources are not scarce either, but they are scattered: a blog post about LoRA, a Colab notebook about QLoRA, a stray discussion about tuning the learning rate. What was missing was one continuous path — something that starts at "what is a tensor" and runs without interruption to "a multi-GPU model trained with DeepSpeed, evaluated, and served with vLLM," while also explaining the real decisions along the way: when you should not fine-tune at all, why RAG is sometimes the better answer, and why your model got worse after training.

This book is that continuous path. I wrote it the way I wish it had been handed to me years ago, when I first trained a 7-billion-parameter model with too little memory and the wrong learning rate, and its output started producing strings that belonged to no nation's language. Out of those experiences — and the hundreds of similar ones I have watched in day-to-day work and in conversations with colleagues — one simple principle has seeped into the structure of the book: every concept is introduced before its tool, every tool before its code, and every piece of code before it is set loose without analysis.