8.9 Fine-tuning: LoRA, QLoRA, PEFT

Work through these

• When fine-tuning beats prompting or RAG — and when it does not

  Fine-tuning teaches form and behaviour; retrieval supplies facts; prompting adjusts both cheaply. Choosing the wrong one is the most common and most expensive mistake in this area.

• LoRA: rank, alpha, target modules

  Training a small number of additional parameters alongside a frozen model, with settings controlling how much capacity is added and where. This is what makes adaptation affordable.

• QLoRA and 4-bit training

  Combining that approach with heavily compressed weights, which brings fine-tuning within reach of a single consumer accelerator. It is the technique that opened this work to individuals.

• Dataset construction and eval before/after

  The dataset is the project: assembling it carefully and measuring the model before and after is what makes fine-tuning a result rather than an anecdote.