Curiosity · Concept

Fine-tuning

Fine-tuning is the process of continuing an LLM's training on your own labeled examples to shift its task format, style, or behaviour. In 2026 most fine-tuning is parameter-efficient — LoRA and its variants update a small adapter rather than full weights — and the decision point is usually 'fine-tune for behaviour, RAG for facts'.

Quick reference

Proficiency
Intermediate
Also known as
PEFT, LoRA, adapter tuning
Prerequisites
RAG basics, Embeddings, Loss functions

Frequently asked questions

What is fine-tuning in LLMs?

Fine-tuning continues training of a pre-trained LLM on your own labeled examples to shift its behaviour — output format, style, tone, or task specialisation. It does not reliably teach new facts (use RAG for that).

Should I fine-tune or use RAG?

Use RAG when the issue is the model doesn't know your facts — private docs, fresh data, anything that changes. Fine-tune when the issue is the model doesn't behave right — wrong format, wrong tone, wrong task structure.

What is LoRA?

LoRA (Low-Rank Adaptation) is a parameter-efficient fine-tuning technique that freezes the base model and trains small low-rank matrices on each attention layer. The resulting adapter is tiny (MBs, not GBs) and can be hot-swapped per task.

Sources

  1. Hu et al. — LoRA paper — accessed 2026-04-20
  2. Hugging Face — PEFT library — accessed 2026-04-20