Fine-Tuning LLMs, Part 1: The Transformer Architecture Guide Nobody Wrote for Fine-Tuners

A complete architectural guide for ML practitioners who need to understand what they are modifying before they modify it. Generated using notebookLM Most engineers who fine-tune language models treat the model as a black box with knobs. They set a rank, pick some target modules from a blog post, run the job, and hope the loss curve looks right. That works, until it doesn’t. Until the model learns nothing despite a clean loss curve, or collapses on day two of