note · growing

Fine-tuning methods

Full SFT, LoRA, QLoRA, DPO, GRPO, distillation — and when each one is the answer.

Training from scratch is almost no product team's job. In practice the choice is which parameter-efficient method.

MethodWhat it doesWhen
Full SFTupdates every weightrarely — costly, forgetful
LoRAlow-rank adapter on a frozen modelthe default starting point
QLoRALoRA over a 4-bit basefitting onto one GPU
DPOaligns on preference pairsyou have good/bad answer pairs
GRPORL against a reward functioncorrectness is programmatically checkable
Distillationcopies a big model into a small onecost and latency pressure
the practical path
Curate200-500 examples
QLoRA SFT
DPOif you have pairs
Evalbefore promoting

An adapter does not replace retrieval. A fine-tuned model still invents the document it never saw — just in a nicer format.

#lora #qlora #dpo

See this note on the whiteboard →