What an adapter is allowed to move
Prompting changes the request. Retrieval changes the documents in the request. Fine-tune changes a slice of weights so the next request, with no extra documents, is more likely to land in the format and label set you trained.
That is why fine-tune is a poor way to teach facts that move. Policy PDFs, inventory, ticket text, last night's incident — those belong in a retriever you can update without a training job. Fine-tune is a decent way to teach a stable output shape: a three-way label, a memo skeleton, a house voice that reviewers already agreed on.
QLoRA (4-bit base + LoRA adapters) is the method this lab names. PEFT holds the adapter math. bitsandbytes holds the 4-bit load. TRL's SFTTrainer runs supervised fine-tune on instruction pairs. None of those libraries are the skill you are unlocking. The skill is deciding the job is narrow enough, the rows are reviewed, and the eval split cannot see the train companies.
Default base: Qwen2.5-1.5B-Instruct. A 0.5B or 3B Instruct sibling in the same family is in-class if VRAM forces the choice. A 32B "just in case" is out of class for this lab. Check the model card for license and terms before you download weights. Do not train on private client logs you do not have rights to keep.
Cost, said plainly: a T4 on Colab, free or quota-limited, often finishes 200–500 short examples in tens of minutes at 1.5B QLoRA. RunPod and Modal are paid escape hatches. You can pass the in-browser gate with a fixture receipt if the config and the judgment payload are honest.