Specialty Digest

Global English editionLearners’ Study Dashboard — in development
The Knowledge Library for All

Fine-Tuning

nounLarge Language Modelsalso model fine-tuning, fine tuning, supervised fine-tuning

In one line

Fine-tuning is the process of further training a pretrained model on a smaller, targeted dataset so that it performs better on a specific task, domain or style.

Fine-tuning is the process of further training a pretrained model on a smaller, targeted dataset so that it performs better on a specific task, domain or style. Instead of building a model from scratch, which needs vast data and computing power, developers start with a model that has already learned general patterns during pretraining and adjust it with a modest amount of specialized examples. A general language model can become a legal contract reviewer, a medical note summarizer or a customer support assistant with the voice of a company. Fine-tuning is a standard form of transfer learning and is used with language, vision, speech and image-generation models.

History of fine-tuning

Fine-tuning became common in computer vision in the early 2010s, when researchers found that taking a network trained on ImageNet, replacing its final layer and training it on a small new dataset gave strong results for tasks such as medical imaging. In language, the technique took off in 2018 with ULMFiT, GPT and BERT, which were pretrained on unlabeled text and then fine-tuned for classification, question answering and other tasks. As models grew to billions of parameters, full fine-tuning became expensive, so researchers developed parameter-efficient methods such as adapters in 2019, prefix tuning in 2021 and LoRA in 2021, followed by QLoRA in 2023, which combined LoRA with quantization so large models could be tuned on a single graphics card. Instruction tuning and reinforcement learning from human feedback are widely used forms of fine-tuning for chat assistants.

How fine-tuning works

The basic steps are as follows.

  1. Choose a base model: pick a pretrained model whose size, license and abilities suit the task.
  2. Prepare data: collect examples, such as prompts and desired responses, labeled documents or images with annotations. Quality matters more than quantity, and a few hundred to a few thousand well-chosen examples can help.
  3. Train: continue training with a small learning rate, so the model adjusts without destroying what it learned. The model updates its weights through backpropagation and gradient descent.
  4. Evaluate: test on held-out data and on general benchmarks to see whether the model improved and whether it lost other abilities.
  5. Deploy and monitor: serve the tuned model and watch performance over time.

Types of fine-tuning

  • Full fine-tuning: all parameters are updated. It can give the best results but needs substantial memory and storage for each tuned copy.
  • Parameter-efficient fine-tuning: only a small set of parameters is trained. Examples include adapters, prompt or prefix tuning and LoRA, which adds small low-rank matrices to existing weights, so the original model stays unchanged and several lightweight adapters can be swapped in.
  • Supervised fine-tuning: training on labeled input and output pairs, the most common form.
  • Instruction tuning: training on diverse instructions and responses to improve following directions.
  • Preference tuning: training with human or AI preferences, as in RLHF and direct preference optimization.
  • Domain adaptation: training on domain text, such as legal or biomedical writing.
  • Fine-tuning of image models: methods such as DreamBooth and LoRA teach diffusion models a particular subject or style from a few images.

Fine-tuning compared with other approaches

Fine-tuning is one way to adapt a model, and others may be cheaper or better depending on the goal. Prompting with clear instructions and examples needs no training and is the first thing to try. Retrieval-augmented generation supplies relevant documents at question time, which is better for facts that change or must be cited. Fine-tuning is best for teaching consistent style, format or behavior, specialized terminology, or skills that are hard to describe in a prompt, and for reducing prompt length and cost at scale. Many systems combine retrieval for knowledge with fine-tuning for behavior.

Applications

Companies fine-tune models for customer support in their tone, for coding assistants tuned on their codebase, for classification of documents or support tickets, for extraction of structured data from invoices, for medical and legal writing assistants and for translation in specialized fields. Creative users tune image models on a personal art style or a product. Researchers adapt models to low-resource languages. Edge developers fine-tune small models to run on devices.

Risks and limitations

  • Catastrophic forgetting: training on narrow data can erode general abilities and safety behavior.
  • Overfitting: a model may memorize a small dataset and perform poorly on new examples.
  • Data quality and bias: errors and biases in the tuning data transfer to the model.
  • Safety degradation: studies have shown that fine-tuning, even on harmless data, can weaken safety training, so tuned models need to be re-tested.
  • Privacy: tuning data may be memorized and leaked, so sensitive data should be handled with care or protected with methods such as differential privacy.
  • Cost and maintenance: each tuned model must be evaluated and updated when the base model changes.

Best practices

Start with prompting and retrieval, define a clear success metric, build a clean and diverse dataset, hold out a test set, begin with a small learning rate and parameter-efficient methods, compare against the base model, test safety and general ability after tuning and document the data and settings. Keep a human review process for high-stakes uses.

A worked example

A law firm wants a model that drafts clause summaries in its house style. It starts with an open language model and gathers 2,000 examples of clauses paired with approved summaries, with client names removed. Using LoRA on one graphics card, it trains for a few hours. The tuned model produces summaries in the right format, but sometimes misses unusual clauses, so the firm adds retrieval of its playbook, keeps lawyers in the loop and re-evaluates each time the base model is updated.

Preparing a fine-tuning dataset

The dataset is usually the most important part. Examples should match the way the model will be used, so a support assistant needs real questions in the format customers write them and answers that reflect approved policy. Teams remove duplicates, fix errors, strip personal data and cover a range of cases, including unusual ones and examples where the right behavior is to decline or ask for clarification. A common practice is to split the data into training, validation and test sets, and to keep a small gold set of examples reviewed by experts for final checks. Synthetic examples written by a stronger model can add volume, but they should be sampled and reviewed by a person to avoid copying that model’s mistakes.

Measuring success

Success should be defined before training. Task metrics such as accuracy, F1 score or the share of outputs that pass a format check suit classification and extraction, while generated text needs human review or a rubric-based evaluation. Teams compare the tuned model with the base model using the same prompts, track regressions on general benchmarks and test safety behavior, including refusals and handling of sensitive requests. If gains are small, better prompts or retrieval may be a cheaper choice.

Future directions

Research aims at tuning methods that need less data, forget less, keep safety behavior intact and let users personalize models on their own devices without sending data to the cloud. Merging several tuned adapters into one model and updating a model continually as new information arrives are also active topics.

Related concepts

Fine-tuning builds on pretraining and is closely related to llm training as a whole, and its results depend on the embedding and tokenization choices of the base model.

FAQs

What is the difference between fine-tuning and training from scratch? Fine-tuning starts from a pretrained model and needs far less data and compute, while training from scratch begins with random weights.

Is fine-tuning better than prompt engineering? Not always. Prompting is cheaper to try, and fine-tuning helps when you need consistent behavior at scale.

What is LoRA? Low-rank adaptation is a method that trains small added matrices instead of all the weights, which makes fine-tuning cheaper and lets one base model serve many tasks.

Key facts

  • Fine-tuning starts from a pretrained model and updates its weights using task-specific data, usually far smaller than the pretraining data.1
  • Parameter-efficient methods such as LoRA update only a small number of added parameters, which sharply reduces memory and cost.2
  • Fine-tuning on narrow data can cause catastrophic forgetting, in which a model loses earlier abilities.3

Related concepts

Quick checkWhat is fine-tuning in machine learning?Show answer

It is further training of an already pretrained model on a smaller dataset for a specific task or domain, so the model keeps its general knowledge while specializing.

Sources

  1. Howard and Ruder, Universal Language Model Fine-tuning for Text Classification, https://arxiv.org/abs/1801.06146
  2. Hu et al., LoRA: Low-Rank Adaptation of Large Language Models, https://arxiv.org/abs/2106.09685
  3. Kirkpatrick et al., Overcoming catastrophic forgetting in neural networks, https://arxiv.org/abs/1612.00796

Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue