Fine-tuning
Continuing a model's training on your own examples so its behavior changes — baked into the model, not supplied at answer time.
When not to use it
- To add knowledge. This is the most expensive misunderstanding in the field. Fine-tuning teaches behaviour and form, not facts. Facts go in the prompt or come from retrieval.
- Before you've exhausted prompting. A good prompt with a few examples solves a surprising share of what people reach for fine-tuning to fix, at zero training cost and no maintenance.
- When your data changes often. Every meaningful update means retraining, re-evaluating, and redeploying. That's a treadmill you have to keep running.
Reach for something else instead
- Few-shot prompting — put three good examples in the prompt. It's free, instant, and shockingly competitive.
- RAG when the real need was current or private information.
- LoRA and other parameter-efficient methods if you do need to fine-tune. Full fine-tuning of a large model is rarely the right first move on cost alone.
Sources & further reading
- Hu et al. (2021), LoRA: Low-Rank Adaptation of Large Language Models — why full fine-tuning is rarely the right first move.
- Howard & Ruder (2018), Universal Language Model Fine-tuning for Text Classification — the transfer-learning recipe that preceded the LLM era.
- Kirkpatrick et al. (2017), Overcoming catastrophic forgetting in neural networks — the failure mode that shows up after launch.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Fine-tuning on too little data and calling the result overfitting. A few hundred well-chosen, consistent examples usually beat thousands of noisy ones.
- Losing general ability while gaining a narrow one. Models can forget how to do everything else — catastrophic forgetting is real and shows up after launch.
- Never building a held-out evaluation set, so "it feels better" is the only evidence the expensive thing worked.
At a glance
FieldLanguage & LLMs
Core ideachange behavior via extra training
Common methodLoRA (parameter-efficient)
Best forstyle, format, narrow skills
Not forfresh, changing facts
DifficultyAdvanced
Flashcards for this concept
Question
—
Answer
—
1 / 4
Often compared with
RAG vs. Fine-tuning — knowledge at answer time vs. behavior changed in the model.