# Fine-Tuning

AI Glossary · Training, Fine-Tuning & Evals
Canonical URL: https://aisat.uk/glossary/fine-tuning
Author: Lucy Gilmore, Assistant General Counsel, Product, Privacy & AI at Sprinklr | Practical Privacy, Product & AI Governance & Defensible Risk Decisions (https://www.linkedin.com/in/lucy-gilmore/)
Breadcrumb: [Glossary](https://aisat.uk/glossary) > [Training & Evals](https://aisat.uk/glossary?category=training) > Fine-Tuning

## TL;DR

> **Fine-tuning** is the process of taking a pre-trained language model and continuing its training on a curated dataset of domain-specific examples to update its weights. It teaches a model consistent output formatting, specialised classification behaviour, or domain tone at lower inference latency than huge prompts.

## How Fine-Tuning works in practice

Pre-training teaches a base model general language and world patterns over trillions of tokens, but leaves it inefficient at narrow enterprise tasks that demand strict formatting or nuanced domain judgement. Fine-tuning adapts that base model using a smaller, high-quality dataset, typically ranging from 500 to 50,000 verified input-output pairs.

In modern workflows, engineers rarely update every parameter of a large model. Instead, they start with Supervised Fine-Tuning (SFT) using parameter-efficient methods such as LoRA or QLoRA, which freeze the base weights and train compact adapter matrices. If the goal is aligning judgement or tone rather than exact targets, teams follow SFT with preference tuning such as Direct Preference Optimisation (DPO) on pairs of chosen and rejected responses, or Reinforcement Learning with Verifiable Rewards (RLVR) when outputs can be scored deterministically.

For instance, a UK logistics platform processes 2 million customs declarations a month. Prompting a frontier model with 6,000 tokens of HS tariff rules cost too much and took three seconds per parcel. By distilling 25,000 verified classifications from the frontier model into an 8-billion-parameter open-weight model via LoRA fine-tuning, they achieved 98.4 per cent agreement at 90 milliseconds latency.

## Why it matters and trade-offs

Fine-tuning is a tool for teaching behaviour, structure, and latency efficiency, not for teaching rapidly changing facts. When you fine-tune a model on a smaller architecture using outputs distilled from a larger teacher model, you can slash inference costs by tenfold and run workloads inside your own VPC without sending sensitive payloads to third-party APIs.

However, premature fine-tuning is one of the most expensive mistakes in applied AI. It creates an ongoing MLOps burden: every base-model upgrade requires re-training and re-evaluating your adapters, and poorly curated training data causes catastrophic forgetting or baked-in hallucinations. Teams should always exhaust context engineering, few-shot prompting, and RAG before committing to a fine-tuning pipeline.

## LinkedIn explainer post

- [Lucy Gilmore on Fine-Tuning (Fine-Tuning)](https://www.linkedin.com/feed/update/urn:li:share:7520930751324930052/)

## Related AI concepts

- [LoRA (Low-Rank Adaptation)](https://aisat.uk/glossary/lora): LoRA (Low-Rank Adaptation) is a parameter-efficient fine-tuning method that freezes a pre-trained model's weight matrices and injects small, trainable rank-decomposition matrices into transformer layers, cutting GPU memory and training cost by over 90%.
- [Small Language Model (SLM)](https://aisat.uk/glossary/small-language-model): A Small Language Model (SLM) is a compact neural language model, typically between one billion and ten billion parameters, optimised for low-latency, low-cost inference on a single GPU or on-device NPU for domain-specific workflows.
- [Retrieval-Augmented Generation (RAG)](https://aisat.uk/glossary/retrieval-augmented-generation): Retrieval-Augmented Generation (RAG) is an AI architecture that retrieves relevant passages from external documents or databases at query time and injects them into a language model's context window to generate cited, grounded answers.
- [Evals (Evaluations)](https://aisat.uk/glossary/evals): Evals (Evaluations) are automated, repeatable testing pipelines that score an AI application's outputs against a golden dataset using deterministic assertions, reference metrics and LLM-as-a-judge rubrics before any prompt or model change ships.