Learn AI /Training and Model Adaptation /Supervised fine-tuning

Supervised fine-tuningSFT

Also called: SFT.

Detail

In five seconds

What it is

A pretrained model is further adapted by showing it input–output examples so its parameters better reproduce the target outputs for a chosen task or style.

In plain words

You take a model that already learned general language patterns and train it further on many examples of 'given this input, produce this output' so it becomes better at a specific task or response style.

An everyday example

A senior editor in Addis Ababa gives a junior editor 2,000 short passages with corrected rewrites. In each case, the junior first drafts a rewrite, then compares it with the editor’s approved version and notes the differences. After many rounds, the junior becomes much better at producing rewrites in that organisation’s preferred style. The junior is not learning every fact in the world; they are adapting their output to match the labelled examples they were shown.

Where the story stops working

A person can reflect on explicit editorial rules and ask questions; a model does not do that in the same way. In supervised fine-tuning, the changes are numerical weight updates based on optimisation, not conscious learning from explanation. Also, an editor may generalise from a few policy principles, while a model’s behaviour depends strongly on the coverage and quality of the examples.

Remember thisSupervised fine-tuning does not build a model from scratch; it reshapes a pretrained one by training on labelled examples of the outputs you want.

Understand it

Why it matters

For non-specialists, supervised fine-tuning matters because many claims about a model being 'trained on our data' actually refer to this specific process. It affects procurement, risk, evaluation and governance. If a vendor says a model was adapted for your task, you need to ask what examples were used, who labelled them, what outputs were treated as correct, and how performance changed after tuning. SFT can improve consistency and task performance, but it can also lock in bias, formatting errors or narrow assumptions from the dataset. It is also a different intervention from prompting, retrieval, preference training or adding fresh documents to a search system, so choosing the wrong method can waste money and create false confidence.

How supervised fine-tuning changes a pretrained model

After viewing this, the reader should be able to describe the inputs, training loop and output of supervised fine-tuning, and distinguish it from pretraining and preference-based methods.

01
1. Start from a base model
Use an already pretrained checkpoint rather than training from scratch.
Draw a rounded rectangle labelled 'Pretrained model' with a smaller tag 'general language patterns'.
02
2. Provide labelled examples
Each training item contains an input and the desired output, such as an instruction and a good answer.
Draw a stack of cards labelled 'input' on the left and 'target output' on the right, connected as pairs.
03
3. Compare prediction with target
For each example, the model produces an output, which is compared with the target to compute loss, an error measure.
Show one pair entering the model, a bubble labelled 'model output', and a comparison box labelled 'loss' between output and target.
04
4. Update parameters repeatedly
Optimisation adjusts the model’s weights over many examples so future outputs more closely match the targets.
Add a circular arrow around the comparison box and model, with a label 'repeat over dataset' and small sliders icon for 'parameter updates'.
05
5. Save the adapted checkpoint
The result is a new model version that is usually better at the tuned task, format or style.
Draw a second rounded rectangle labelled 'SFT checkpoint' with tags like 'better instruction format' and 'task-specific behaviour'.

Show a left-to-right sequence of four panels. Start with a general pretrained model, then a labelled dataset of input–target pairs, then repeated comparison between model output and target with parameter updates, and end with an adapted checkpoint specialised for a task. Read from left to right as a process.

Hovering or tapping each panel reveals one sentence on what enters, what is compared, and what changes; a toggle switches labels between 'plain language' and 'technical terms'.

Describe this without the diagram

This diagram presents supervised fine-tuning as a five-step process. First, you begin with a pretrained base model that already captures broad language patterns. Second, you prepare labelled examples, where each item has an input and a target output. Third, for each example, the model generates an output and that output is compared with the target to measure error, called loss. Fourth, the training system updates the model’s parameters repeatedly across many examples to reduce that loss. Fifth, you save the result as a new checkpoint or model version that is more specialised for the target task or format.

With reduced motion: A static five-panel process diagram replaces any looping animation. Arrows show a single left-to-right path from base model, to labelled examples, to comparison with target, to repeated updates noted by a text label, to the final adapted checkpoint.

Supervised fine-tuning takes an existing model and adjusts it using labelled examples. A labelled example means there is an input and a target output: for example, a passage of text and the correct answer, summary or rewrite. The word supervised means the model is trained against known targets, not left to guess what good behaviour looks like only from comparison or rewards.

What goes in is usually a dataset of many prompt–response pairs or task examples. During training, the model reads the input, produces a prediction, compares that prediction with the target output, and calculates an error signal called a loss. The training system then updates the model’s parameters, which are the numerical weights that shape its output behaviour.

What happens overall is not the creation of a new model from nothing, but the adaptation of a pretrained one. Pretraining gives broad language patterns. Supervised fine-tuning pushes the model toward a narrower goal, such as following an instruction format, extracting fields, classifying text, or answering in a house style. The result is a new checkpoint or model version that is usually better on the examples it was tuned for.

This is often confused with instruction tuning and with reinforcement learning from human feedback. Instruction tuning is a common kind of supervised fine-tuning focused on following instructions. Reinforcement learning from human feedback uses a different training signal, usually rewards derived from preferences, rather than direct target outputs for each example.

The quality of supervised fine-tuning depends heavily on the dataset. If the examples are noisy, biased, too narrow or badly formatted, the model may learn those problems too. It also does not reliably add fresh factual knowledge in a controlled way, so it should not be treated as a secure way to update facts or policy content without evaluation.

How “Marking examples to teach a consistent style” maps onto the real thing

In the story	In the AI system
approved rewrites paired with original passages	labelled input–output training examples
comparing the draft with the approved version	calculating loss from the difference between prediction and target
many rounds of correction	iterative parameter updates during fine-tuning
better matching the organisation’s style	adapting a pretrained model to a narrower task or format

A worked example

A programme officer in a regional bureau is testing whether a pretrained model can turn approved monitoring extracts into first drafts of results paragraphs for a progress report before the quarterly clearance round.

Starting from: Example pair: Input: an approved monitoring extract plus the instruction, "Draft the results paragraph." Target output: the corresponding cleared paragraph in the office's standard format.

Start from a pretrained checkpoint rather than training a model from scratch.
Assemble approved, non-sensitive input-output pairs and reserve a separate set for evaluation. Each input is a monitoring extract and instruction; each target is the cleared paragraph.
For each training pair, have the model predict the target sequence token by token. Calculate cross-entropy loss between its predicted token distributions and the target tokens.
Use gradient-based updates across successive batches to adjust some or all model parameters, producing a new checkpoint.
Compare the original and tuned checkpoints on the reserved examples, checking format adherence, factual consistency and performance across the required languages.

Result: The tuned checkpoint is more likely to reproduce the patterns represented in the cleared examples. It may still omit, distort or invent content, particularly when an input differs from the training data, so its paragraphs remain drafts for review and clearance.

In a United Nations workplace

At close of business, a chief of section in a Resident Coordinator's Office must clear talking points requested by Headquarters for informals the next morning. The office is considering adapting a pretrained model using pairs of cleared background notes and final talking points, so first drafts follow the established structure and tone.

Supervised fine-tuning would use those input-output pairs to shift the model towards the office's preferred format, phrasing and level of detail. It can improve consistency, but it does not make the checkpoint an authoritative source for facts or replace clearance.

Before you try thisThe training set must contain only material approved for this use and be processed in an approved environment. Confidential, personal, protection-sensitive or security-sensitive material must not be entered into an external or otherwise unapproved tool, because fine-tuning can expose or reproduce training content. Every draft still requires substantive review and clearance.

Where you may hear it

At an ICT governance board reviewing a model-adaptation proposal
During a vendor demo in a procurement meeting
In an evaluation office discussion on automated report drafting
At a regional bureau review of a training and validation plan

Avoid confusion

Often confused with

Common misconceptions

Common misconception: Supervised fine-tuning simply feeds the model our documents so it 'knows' them.
Correction: Not usually. SFT uses labelled input–output examples and changes model parameters to better match those targets. A pile of raw documents without target outputs is not, by itself, supervised fine-tuning. Even when documents are turned into examples, the result is behavioural adaptation, not a guaranteed or auditable store of every fact in those documents.
Common misconception: Instruction tuning and supervised fine-tuning are different unrelated methods.
Correction: Instruction tuning is usually a subtype of supervised fine-tuning. The difference is scope: SFT is the broader category of training on labelled examples, while instruction tuning focuses on examples where the model learns to respond to instructions in a desired format or style.
Common misconception: If we do supervised fine-tuning, the model will become more truthful overall.
Correction: It may improve on the kinds of examples it was trained on, but truthfulness does not automatically follow. If targets are incomplete, oversimplified or inconsistent, the model can learn those weaknesses. Reliability still needs careful evaluation on representative test cases.
Common misconception: Supervised fine-tuning and RLHF are basically the same because both use human data.
Correction: Both can involve human-created data, but the training signal differs. In SFT, the model is trained directly against target outputs for each example. In RLHF, a reward model or preference signal guides optimisation toward preferred outputs rather than one fixed target answer.

Quick check

Which description best matches supervised fine-tuning?

Go deeper

The simple version

The technical version

Supervised fine-tuning starts from a pretrained checkpoint and continues optimisation on a labelled dataset D={(x,y)}. For each input x, the model defines a probability distribution over output tokens or labels. Training minimises a supervised objective, commonly cross-entropy loss between the model’s predicted distribution and the target sequence y, using gradient-based updates to some or all parameters. The output is a new checkpoint whose conditional behaviour P(y|x) has been shifted toward the task distribution represented in the dataset. In practice, SFT may update all weights or only a subset through parameter-efficient methods such as LoRA. It is distinct from pretraining because the data are labelled for a downstream objective, and distinct from preference or reinforcement-based methods because the supervision is direct targets rather than scalar rewards or pairwise preferences.

Reviewed external resources

EncyclopaediaWikipedia search: Supervised fine-tuning (opens in a new tab)Wikipedia
This Wikipedia search results page helps a reader find articles related to supervised fine-tuning and suits someone who wants broad starting points rather than one curated source.
search results, not one pagelink checked 2026-07-31
CourseLLM Course (opens in a new tab)Hugging Face
This Hugging Face course gives structured lessons on how large language model adaptation works and suits readers who want a guided technical introduction.
Deeperlink checked 2026-07-31
CourseShort Courses (opens in a new tab)DeepLearning.AI
These DeepLearning.AI short courses offer practical learning paths on model adaptation workflows and suit readers who prefer applied, instructor-led explanations.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Supervised fine-tuning (opens in a new tab)YouTube search
This YouTube search results page helps readers find a range of video explanations of supervised fine-tuning and suits people who learn best by watching, but it is not one vetted lesson.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Supervised fine-tuning (opens in a new tab)Reddit search
This Reddit search results page shows community discussions and practical questions about supervised fine-tuning and suits readers who want to see common confusions and practitioner experiences, with caution about reliability.
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

Supervised fine-tuning does not build a model from scratch; it reshapes a pretrained one by training on labelled examples of the outputs you want.