🤖 AI / ML

Fine-tuning

fine-tuning · fine tuning · finetuning · model fine-tuning · transfer learning · instruction tuning

Definition

Fine-tuning is a transfer learning method in which a pre-trained model continues optimising on a specialised target data set in order to adapt to a particular task or domain.

What fine-tuning is

A pre-trained model — an LLM, an embedding encoder, a vision model — has been trained on general
data and carries broad knowledge. Fine-tuning is a second training stage on a specialised sample.
The aim is to adapt the model to a specific task, domain or style without training from scratch.

It is far cheaper in resources: most of the knowledge is already baked into the weights from
pre-training, and all that remains is to nudge them toward the target domain.

Varieties of fine-tuning

Approach	Description	When to use it
Full fine-tuning	All the model’s weights are updated	Plenty of data, deep specialisation
LoRA / PEFT	A small fraction of weights is updated	Limited compute budget
Instruction tuning	Training on instruction-response pairs	Changing behaviour and style
RLHF	Reinforcement learning from human raters	Aligning with human preference

Fine-tuning in e-commerce

The main retail applications:

AI shopping assistant. Fine-tuning on historical shopper conversations and product
documentation lets the model use the right terminology, hold the brand’s tone of voice and return
fewer irrelevant answers.

Product embeddings. Fine-tuning multilingual encoders on pairs of query and relevant product
improves semantic search and recommendation quality in specific categories such as fashion, DIY
and jewellery.

Intent classification. Tuning small classifiers to the intent categories of one particular store.

Important: fine-tuning is not a replacement for RAG when data keeps changing. The catalogue
moves every day; sewing it into model weights makes no sense. RAG refreshes knowledge in real
time, while fine-tuning changes the model’s behaviour and style.

Common mistakes

Catastrophic forgetting: training on a narrow data set can displace the model’s general
knowledge. The fix is PEFT, or mixing target data with a share of general data.
Overfitting on a small sample: a tiny data set plus many epochs gives an excellent training
score and poor answers on real queries. Watch the metrics on a validation set.
Data quality beats data volume: noisy or contradictory training examples make the model worse.
Invest in annotation, not in quantity.

Frequently asked questions

How is fine-tuning different from RAG?

RAG (retrieval-augmented generation) enriches a request with relevant documents at runtime, leaving the model itself unchanged. Fine-tuning changes the model's weights through further training. RAG is simpler and cheaper for knowledge that keeps updating, such as a product catalogue; fine-tuning is more effective for changing the style of answers or for narrow specialised tasks.

How much data does fine-tuning an LLM need?

It depends on the goal. For adapting style and tone of voice (instruction fine-tuning), 500 to 5,000 question-and-answer examples are usually enough. Domain knowledge needs thousands of examples. The decisive condition is data quality: 1,000 clean examples beat 10,000 noisy ones.

What are PEFT and LoRA?

Parameter-efficient fine-tuning (PEFT) is the family of methods that update only a small fraction of a model's weights. LoRA (low-rank adaptation) is the most common one: it adds low-rank matrices to the model's layers. This cuts compute cost by a factor of 10 to 100 compared with full fine-tuning.

When do you not need fine-tuning?

If the task is solved by a well-written prompt or by RAG, fine-tuning adds nothing. It is justified when you need to change the model's baseline behaviour, specialise it for a narrow domain, or cut inference cost by moving to a smaller model.