Fine-Tuning — AI Glossary

Reference details

Related terms

• LoRA (Low-Rank Adaptation)
• RLHF (Reinforcement Learning from Human Feedback)
• Supervised Fine-Tuning (SFT)
• Synthetic Data
• Parameter-Efficient Fine-Tuning (PEFT)

How it works

1. A curated set is formatted, a loss is computed on the target tokens, and gradients are applied at a low learning rate — the mechanics are ordinary supervised learning on a model that already works.
2. Full fine-tuning updates every weight, so it needs optimiser state for every parameter, which is the term that puts it out of reach rather than the weights themselves.
3. Parameter-efficient methods freeze the model and train a small attached set instead, so optimiser state scales with the addition and a single accelerator becomes sufficient.
4. Both produce a new artifact that has to be versioned, served, and re-evaluated, which means the recurring cost is the lifecycle rather than the training run.
5. Because the result is weights, a mistake is not revertible in the way a prompt change is: rolling back means serving the previous checkpoint, which has to still exist.
6. That irreversibility is the practical argument for exhausting prompting and retrieval first — both are edits to a request, and this is an edit to the thing serving every request.

Example: The question that decides it

A model that will not hold your house JSON format across 50 turns needs fine-tuning: the behaviour is the problem. A model that does not know your Q3 pricing needs retrieval: the knowledge is the problem. Fine-tuning teaches form and style reliably and is a poor and expensive way to install facts, because facts change and weights do not.

Commonly confused with

Fine-tuning adjusts behaviour far more reliably than it adds knowledge. Facts trained into weights cannot be cited, updated or revoked per tenant, and they go stale silently — which is why retrieval, not fine-tuning, is the normal answer to “the model does not know X”.

When to use it

Reach for it when:

• You need a consistent format, tone or structure that prompting keeps missing
• The task is narrow and you have curated examples of it done correctly
• Prompt length has become a recurring cost you want to compile away

Reach for something else when:

• The knowledge changes — you will be retraining forever
• You have a few hundred examples; that usually teaches noise
• You have not yet exhausted prompting and retrieval, which are cheaper to reverse

Referenced by

This term is used by the following ModelRefs references:

• lora fine tuning stack
• full fine tuning stack
• learn fine tuning
• become llm engineer

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Fine-Tuning — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path