Learn AI /Training and Model Adaptation /Fine-tuning
Fine-tuning
- Starter
- Training and Model Adaptation
- Policy officers
- Managers & front offices
- Technical readers
Also called: finetuning, fine tune.
Back to the milestone you came from
In five seconds
What it is
Updating a pre-trained model's parameters with additional examples to improve performance on a narrower task, style or domain.
In plain words
Fine-tuning means taking a general AI model and giving it extra practice on a narrower job so it performs that job more consistently.
An everyday example
A ready-made jacket already has a usable shape. A tailor then uses several fittings to adjust its seams, sleeves and waist for a particular wearer. The fittings do not create the jacket from scratch, but they change the jacket itself. Fine-tuning works similarly: a pre-trained model is adjusted through further training so that its behaviour better fits a narrower task. Poor measurements can produce a poor fit, just as weak or unrepresentative examples can produce an unreliable adapted model.
A human translator can consciously ask questions, justify choices and transfer lessons between topics in ways a model cannot reliably do. In fine-tuning, the system is not reading examples with human awareness; numerical parameters are being adjusted. Also, a human can often notice a bad house style and resist it, while a model may simply absorb harmful patterns from the fine-tuning data.
Remember thisFine-tuning updates a pre-trained model's parameters using additional examples and produces a new version for a narrower job. It is not prompting, and it still requires authorised data, evaluation and human review.
This term is contested
People use 'fine-tuning' in both a narrow and a broad sense. Some use it only for parameter-updating training on a pre-trained model, while others use it loosely for many kinds of model adaptation, including parameter-efficient methods.
Understand it
Understand it
Why it matters
Fine-tuning matters because organisations often need more than a general model. They may want a model that follows a specific format, works better in a specialist domain, or uses approved terminology more consistently. For leaders and policy staff, this affects cost, risk, procurement and governance: what data can be used, who checks quality, how a new model version is evaluated, and whether changing the model is even necessary. In many cases, a prompt or document retrieval system is enough. In other cases, fine-tuning can improve reliability on a narrow task. Knowing the difference helps you ask better questions before spending money or exposing data.
How fine-tuning changes a general model into a task-specific version
After viewing this, the reader should be able to explain that fine-tuning starts with an existing model, adds task-specific training data, updates parameters, and produces a new version better suited to a narrower use.
01
1. Start with a base model
Begin with a model already trained on broad data. It has general capabilities but may not match the exact task, format or terminology you need.
Draw a large rounded rectangle labelled 'Base model'. Add a small caption below: 'General capabilities'.
02
2. Prepare task-specific examples
Collect and clean examples that represent the target task, such as prompts and good answers, translations, classifications or summaries.
Place a stack-of-documents icon in a second box labelled 'Fine-tuning data'. Add small tags such as 'examples', 'labels', 'quality matters'.
03
3. Run additional training
Use the examples to adjust the model’s parameters. This strengthens patterns useful for the target task. Some methods change many parameters; others change only a small subset.
Show a gear or slider icon in a third box labelled 'Parameter updates'. Add a side note: 'Not training from scratch'.
04
4. Save a new model version
The result is a new checkpoint or model version that usually performs better on the narrower task, though it may also become less general or copy biases in the data.
Draw a final box labelled 'Fine-tuned model'. Add two small callouts: 'Better on target task' and 'Still needs testing and oversight'.
A left-to-right flow with four boxes connected by arrows. Read from the general starting model, through curated examples and an update step, to a new model version for a narrower task. Small side notes warn that data quality affects behaviour and that the output remains a model, not a guarantee of correctness.
Hovering or tapping each box reveals a short note with one example: translation, summary, classification or style adaptation.
Describe this without the diagram
The diagram is a four-step flow read from left to right. First is a box called Base model, described as a general model with broad capabilities. An arrow leads to Fine-tuning data, shown as curated task-specific examples where data quality matters. Another arrow leads to Parameter updates, explaining that additional training changes the model’s numeric settings and that this is not training from scratch. A final arrow leads to Fine-tuned model, a new saved version that is usually better on the target task but still needs evaluation, because fine-tuning can also narrow behaviour or reproduce errors in the data.
With reduced motion: A static four-box flow replaces animation. Arrows remain visible between Base model, Fine-tuning data, Parameter updates and Fine-tuned model, with the same labels and caution notes shown at all times.
A large model usually starts as a general system trained on very broad data. Fine-tuning is a later adaptation step. You take that existing model, show it more task-specific examples, and update its internal weights, meaning the numeric settings that shape its outputs.
What goes in is usually a starting model plus a curated dataset. The dataset might contain input-output pairs, such as a customer question and a good answer, or examples from a specialist domain such as law, health or agriculture. The quality and labelling of these examples matters a lot.
What happens is additional training, but not from scratch. The model is adjusted so patterns useful for the new task become stronger. Depending on the method, this may update all weights or only a small subset. The result is a new model version, often called a fine-tuned checkpoint.
What comes out is not guaranteed truth or expertise. A fine-tuned model may become more consistent, better at following a format, or better with a domain vocabulary. But it can also become narrower, inherit errors from the new data, or perform worse outside the examples it was adapted on.
People often confuse fine-tuning with prompt engineering or retrieval-augmented generation. Prompt engineering changes the instructions you give at use time. Retrieval-augmented generation gives the model external documents at use time. Fine-tuning changes the model itself before use.
How “Tailoring a ready-made jacket” maps onto the real thing
| In the story | In the AI system |
|---|---|
| the skilled general translator | the pre-trained base model |
| past field reports with preferred translations | the fine-tuning dataset |
| practising on those examples | additional training that updates model weights |
| more consistent organisational style | better performance on a specific task or domain |
A worked example
During a first reading in a General Assembly Main Committee, a political affairs desk officer must sort paragraphs from public statements under the headings of the 6pm briefing note. The general model applies the headings unevenly, so an information management officer tests fine-tuning on paragraph-and-label pairs previously reviewed for that workflow.
Starting from: An approved pre-trained model and a dataset of public statement paragraphs with labels checked by desk officers.
- The information management officer removes duplicates, reconciles inconsistent labels and divides the reviewed examples into training and evaluation sets.
- Training continues from the pre-trained checkpoint. Gradient-based updates modify selected model parameters so that the reviewed classification patterns become more likely.
- The resulting model version is tested on held-out paragraphs that were not used for parameter updates, and its results are compared with those of the original model.
- The desk officer examines missed themes, overlapping headings and passages that depend on surrounding context before deciding whether the adapted model is suitable for the briefing workflow.
Result: The fine-tuned version uses the established headings more consistently and reduces initial sorting. It still cannot judge the negotiating significance of a paragraph or reliably resolve ambiguous language. The desk officer remains responsible for the briefing note and its clearance.
In a United Nations workplace
At a small field presence with intermittent connectivity, a national programme officer has a queue of monitoring-report extracts awaiting French translation. The monthly progress report is due to the regional bureau by close of business. The office is considering fine-tuning an approved pre-trained model on reviewed source-and-translation pairs, then running the adapted version in an approved local environment.
Fine-tuning would continue training from the existing model checkpoint. Parameter updates would make the terminology, structure and translation choices in the reviewed examples more likely in future outputs. The result would be a distinct model version for evaluation, not merely a longer prompt or a reference file attached to each request.
Before you try thisRaw partner submissions, names, contact details, protection information and unpublished operational text must not enter a fine-tuning pipeline unless both the tool and the specific use are approved for that material. Training examples may be memorised and reproduced. Minimise and de-identify the dataset, confirm permissions, test for leakage and translation errors, and retain human review and clearance.
Where you may hear it
- At an ICT governance board considering whether an adapted model may move from pilot to production
- In a procurement clarification meeting after a bidder proposes fine-tuning on the office's examples
- During acceptance testing, when a programme officer asks which parameters changed and how the new version was evaluated
- At a data protection review before approved records are assembled into a training dataset
Avoid confusion
Avoid confusion
Often confused with
Prompt engineering
The practice of writing and refining instructions so a language model is more likely to produce a useful, safe and well-structured reply.
Retrieval-augmented generation
A method that first finds relevant source passages and then asks a language model to answer using that material.
Training
Training is the process of adjusting a model's parameters using data or feedback to improve performance on a learning objective.
In-context learning
A model uses examples and instructions placed in the current prompt to imitate a pattern for that one interaction, without changing its stored parameters.
Side-by-side
Prompt engineering vs fine-tuning
Both are called "customising the model". Only one changes it.
RAG vs fine-tuning
We want it to know our documents. Which one do we need?
Common misconceptions
Common misconception: Fine-tuning simply means the model has read our documents.
Correction: Not exactly. Fine-tuning is a training process that changes model parameters using curated examples. Giving documents at use time is different; that is closer to retrieval-augmented generation or context prompting, where the model itself may remain unchanged.
Common misconception: Any fine-tuned model will be smarter than the original one.
Correction: A fine-tuned model is usually more specialised, not universally better. It may improve on the target task while becoming less flexible elsewhere, and poor-quality data can make behaviour worse rather than better.
Common misconception: If we fine-tune on internal material, the output becomes authoritative.
Correction: No. Fine-tuning can make style and domain language more consistent, but it does not guarantee factual accuracy, policy compliance or legal correctness. Evaluation and human oversight are still needed.
Common misconception: Fine-tuning and prompt engineering are the same thing.
Correction: They happen at different stages. Prompt engineering changes the instructions or examples you provide when using a model. Fine-tuning changes the model beforehand by updating parameters through additional training.
Quick check
Which situation is the clearest example of fine-tuning?
Fine-tuning means additional training on a starting model using task-specific examples, which changes the model’s parameters. The other options describe prompt engineering, retrieval-augmented generation and model compression.
Go deeper
Go deeper
The simple version
Fine-tuning means taking a general AI model and giving it extra practice on a narrower job so it performs that job more consistently.
The technical version
Fine-tuning is post-pretraining optimisation of an existing model on a target dataset. Starting from a pre-trained checkpoint, training continues on examples such as prompt-response pairs, labelled classifications or preferred outputs. Gradient-based updates modify model parameters, either across the full network or through parameter-efficient methods such as LoRA, to adapt behaviour to a domain, task, format or instruction style. The output is a new checkpoint or model version. This differs from in-context learning, where the model stays unchanged and only the prompt context varies at inference time, and from retrieval-augmented generation, where external documents are supplied at inference time without necessarily altering parameters.
Reviewed external resources
EncyclopaediaFine-tuning (opens in a new tab)Wikipedia
This link gives a broad introductory overview of fine-tuning and suits readers who want a quick general reference before going deeper.
Intermediatelink checked 2026-07-31
CourseLLM Course (opens in a new tab)Hugging Face
This link gives a structured course-style explanation of how large language models are adapted and suits learners who want technical depth step by step.
Deeperlink checked 2026-07-31
CourseShort Courses (opens in a new tab)DeepLearning.AI
This link gives short practical lessons on fine-tuning-related workflows and suits busy professionals who prefer brief guided instruction.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Fine-tuning (opens in a new tab)YouTube search
This link is a search results page for video explanations of fine-tuning and suits readers who learn best by comparing multiple visual explainers.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Fine-tuning (opens in a new tab)Reddit search
This link is a search results page for community discussions about fine-tuning and suits readers who want to see common questions, experiences and informal debate.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Fine-tuning updates a pre-trained model's parameters using additional examples and produces a new version for a narrower job. It is not prompting, and it still requires authorised data, evaluation and human review.