Fine-tuning
A base model is extraordinarily knowledgeable but has no concept of following instructions — it just continues text. Fine-tuning trains it on instruction-response pairs to redirect that knowledge into helpful behavior. LoRA does this by updating just 0.006% of the weights. The knowledge was always there.
It won’t answer
A model fresh from pre-training does this. It doesn’t answer. It continues.
It has seen lists of exercises like this online, so it carries on the list.
Guess what follows
This model reads “What is the capital of France?” Which word does it put next?
Show it answers
Fine-tuning trains it further on examples of a request followed by a good reply.
The same loop as pre-training, on a far smaller set of carefully written examples.
Before and after
Same question, same knowledge. What changes is what it does with your text.
Behaviour, not facts
Mostly not. Fine-tuning teaches it how to use what it already learned.
The French was already in there from pre-training. Fine-tuning taught it to hand that over as an answer.
Prompt or fine-tune
Is a good prompt enough, or is this worth fine-tuning?
Tap a box for each job.
Back to bonjour
The same model, after fine-tuning, finally answers the question.
From continuer to assistant
Fine-tuning trains a base model on example replies, so it learns to answer.
The model before this step is the base model. Training on example replies is called fine-tuning.