Start a project
Techniques Reviewed 24 Sep 2026 · 15 sources

fine-tuning

/ˈfaɪnˈtuː.nɪŋ/ noun Verb fine-tune · adjective fine-tuned · also spelled finetuning
Definition

A technique that continues training an already trained AI model on a smaller set of examples, adjusting its weights so it suits a particular task, subject or style.

Examples
  • GPT-3 trained further on answers written by human labelers
  • An image network trained on one task, retrained to detect objects
  • GPT-3.5 Turbo trained to write in a company’s brand voice
A model stands beside a huge, faded pile of pages it has already read. A thin stack of example cards feeds into it, and a few of the many dials on its side turn slightly.
Inkling · A trained model nudged by a thin stack of examplesFig. F1
Earliest AI use
1988
Speech recognition: extra training for a net built from trained parts.
Current meaning
2013
Vision: a network trained for one task, adapted to another with few labels.
Opposite
From scratch
Training from random weights, with no trained model to start from.
§01

In plain words

Fine-tuning is further training of an already trained AI model on a smaller set of examples, so that it suits a particular task, subject or style.1

A large language model first goes through pretraining, in which it learns general patterns from a vast collection of text. Fine-tuning is a later and much smaller round of training, on examples chosen for one job. It adjusts the model’s weights, the numbers that store what it has learned, so the new behavior stays in the model.

Companies use it for behavior that is hard to get from instructions alone, such as a consistent output format or a brand’s tone. OpenAI opened fine-tuning for GPT-3.5 Turbo in 2023. It said early testers had cut their prompts by up to 90% by training the instructions into the model.2 A fine-tuned model can also beat a far larger one. Human evaluators preferred answers from a 1.3-billion-parameter InstructGPT, trained further on human examples and feedback, to those of the 175-billion-parameter GPT-3.3

Fine-tuning is often confused with retrieval-augmented generation, which leaves the weights alone and supplies relevant documents with each question. Prompt engineering also leaves the model unchanged. It rewrites only the text sent to the model.

The extra training can have side effects. In 2023, researchers undid GPT-3.5 Turbo’s safety training by fine-tuning it on 10 examples written as an attack, at a cost of under $0.20. Harmless training data weakened it too, though less.4 In May 2026, OpenAI began winding down its self-serve fine-tuning service. Organizations that had never used it can no longer train models, and active customers can start new training jobs until 6 January 2027. Models already fine-tuned keep running until their base models are retired.

§02

How it works

Fine-tuning runs the same learning process as pretraining, from a different starting point. The model begins with its pretrained weights, not random ones. Each training example pairs an input with the output wanted, such as a customer question and an ideal reply. The model writes its own answer, the gap from the target is measured, and the weights shift slightly to narrow it.

Because the model already knows the language, the data can be small. OpenAI said in 2021 that fine-tuning GPT-3 began to show benefits with fewer than 100 examples.5 In 2018, the ULMFiT method, given only 100 labeled examples, matched a model trained from scratch on 100 times as much data.

Full fine-tuning updates every weight, so each task needs its own copy of a large model. Parameter-efficient methods train far less. LoRA freezes the original weights and trains small added matrices. On GPT-3, it cut the trainable parameters 10,000-fold and the GPU memory needed threefold.6

OpenAI’s InstructGPT was fine-tuned in two stages. Supervised fine-tuning first trained GPT-3 on example answers written by labelers. Reinforcement learning from human feedback then trained it further on labelers’ rankings of its own answers.

Narrow training can have broad effects. In a 2025 study, a model fine-tuned to write insecure code began asserting, on prompts unrelated to coding, that humans should be enslaved by AI.

§03

Origin and first known use

Formed from fine and tune. The Online Etymology Dictionary dates fine-tuning to 1909, in radio, and the verb fine-tune to 1969. Merriam-Webster dates the verb to 1959. In AI, Alex Waibel wrote of “all-net fine tuning” in 1988, for small extra adjustments to a speech network assembled from trained parts. In 2013, image researchers used the term for adapting a network trained on one task to a new one, the sense in use today. From 2018 it spread to language models. In 2021 OpenAI began offering it to developers. The one-word spelling finetuning also appears.

  1. 1909 The Online Etymology Dictionary dates fine-tuning to 1909, in reference to radio, and the verb fine-tune to 1969, formed from the noun.
    Online Etymology Dictionary, “fine-tune” etymonline.com →: source for 1909
  2. 1988 Alex Waibel assembles a large speech recognition network from smaller trained ones, then frees all its connections for small extra adjustments, a step he calls “all-net fine tuning”.
    Waibel, “Consonant Recognition by Modular Construction of Large Phonemic Time-Delay Neural Networks”, Advances in Neural Information Processing Systems 1 (NIPS 1988) papers.nips.cc →: source for 1988
  3. 2013 Girshick and colleagues take an image network trained on one task and fine-tune it to detect objects. When labeled data is scarce, this gives a significant boost.
    Girshick, Donahue, Darrell and Malik, “Rich feature hierarchies for accurate object detection and semantic segmentation”, CVPR 2014 arXiv 1311.2524 →: source for 2013
  4. 2018 Howard and Ruder’s ULMFiT fine-tunes a pretrained language model to classify text. With 100 labeled examples, it matches a model trained from scratch on 100 times more data.
    Howard and Ruder, “Universal Language Model Fine-tuning for Text Classification”, ACL 2018 arXiv 1801.06146 →: source for 2018
  5. 2021 Hu and colleagues propose LoRA, which freezes a pretrained model’s weights and trains small added matrices instead, so each task needs far fewer trainable parameters.
    Hu et al., “LoRA: Low-Rank Adaptation of Large Language Models” arXiv 2106.09685 →: source for 2021
  6. 2022 OpenAI fine-tunes GPT-3 on answers written by human labelers, then trains it further on their rankings with reinforcement learning. The result is InstructGPT.
    Ouyang et al., “Training language models to follow instructions with human feedback” arXiv 2203.02155 →: source for 2022
  7. 2023 Qi and colleagues undo GPT-3.5 Turbo’s safety alignment by fine-tuning it through OpenAI’s API on 10 adversarially designed examples, at a cost of under $0.20.
    Qi et al., “Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!” arXiv 2310.03693 →: source for 2023
  8. 2026 OpenAI starts winding down self-serve fine-tuning. Organizations new to it can no longer train models; existing fine-tuned models keep running until their base models are retired.
    OpenAI, “Update to OpenAI’s self-serve fine-tuning”, API deprecations page, 7 May 2026 developers.openai.com →: source for 2026
§04

In the wild

9 sightings · 2013–2026 · oldest first
  1. Paper11 Nov 2013

    “…when labeled training data is scarce, supervised pre-training for an auxiliary task, followed by domain-specific fine-tuning, yields a significant performance boost.”

    Girshick et al., CVPR 2014 arXiv 1311.2524 →: Girshick et al., CVPR 2014, 11 Nov 2013Archived →: copy of Girshick et al., CVPR 2014, 11 Nov 2013
  2. Paper18 Jan 2018

    “We propose Universal Language Model Fine-tuning (ULMFiT), an effective transfer learning method that can be applied to any task in NLP…”

    Howard and Ruder, ACL 2018 arXiv 1801.06146 →: Howard and Ruder, ACL 2018, 18 Jan 2018Archived →: copy of Howard and Ruder, ACL 2018, 18 Jan 2018
  3. Paper17 Jun 2021

    “As we pre-train larger models, full fine-tuning, which retrains all model parameters, becomes less feasible.”

    Hu et al., arXiv preprint arXiv 2106.09685 →: Hu et al., arXiv preprint, 17 Jun 2021Archived →: copy of Hu et al., arXiv preprint, 17 Jun 2021
  4. Blog14 Dec 2021

    “Developers can now fine-tune GPT‑3 on their own data, creating a custom version tailored to their application.”

    OpenAI, product announcement openai.com →: OpenAI, product announcement, 14 Dec 2021Archived →: copy of OpenAI, product announcement, 14 Dec 2021
  5. Paper4 Mar 2022

    “…we collect a dataset of labeler demonstrations of the desired model behavior, which we use to fine-tune GPT-3 using supervised learning.”

    Ouyang et al., arXiv preprint arXiv 2203.02155 →: Ouyang et al., arXiv preprint, 4 Mar 2022Archived →: copy of Ouyang et al., arXiv preprint, 4 Mar 2022
  6. Headline22 Aug 2023

    “OpenAI brings fine-tuning to GPT-3.5 Turbo”

    TechCrunch techcrunch.com →: TechCrunch, 22 Aug 2023Archived →: copy of TechCrunch, 22 Aug 2023
  7. Blog10 Jul 2024

    “Fine-tuning is a popular technique to improve model performance.”

    Anthropic, product announcement claude.com →: Anthropic, product announcement, 10 Jul 2024Archived →: copy of Anthropic, product announcement, 10 Jul 2024
  8. Paper24 Feb 2025

    “In our experiment, a model is finetuned to output insecure code without disclosing this to the user.”

    Betley et al., ICML 2025 arXiv 2502.17424 →: Betley et al., ICML 2025, 24 Feb 2025Archived →: copy of Betley et al., ICML 2025, 24 Feb 2025
  9. Docs7 May 2026

    “Creating fine-tuning jobs or training is not available to organizations that have not previously run fine-tuning.”

    OpenAI, API deprecations page developers.openai.com →: OpenAI, API deprecations page, 7 May 2026Archived →: copy of OpenAI, API deprecations page, 7 May 2026
§05

Not to be confused with

  • Pretraining

    Pretraining is the first and costliest stage, building general ability from a huge, mostly unlabeled collection of text. Fine-tuning comes after it, starting from those weights and training on a much smaller, often labeled, set.

    See pretraining →

  • Retrieval-augmented generation

    Fine-tuning changes a model’s weights through further training, so what it learns stays until it is trained again. RAG leaves the weights alone and supplies relevant documents with each question.

    See Retrieval-Augmented Generation →

  • Prompt engineering

    Prompt engineering changes only the text sent with each request and leaves the model as it was. Fine-tuning trains the model on examples and changes its weights, so the new behavior holds without long instructions.

  • RLHF

    Reinforcement learning from human feedback is itself a form of fine-tuning. It learns from people’s rankings of a model’s answers, not from example answers. InstructGPT used supervised fine-tuning first, then RLHF.

    See reinforcement learning from human feedback →
§06

Questions

  • What is fine-tuning in AI?

    Fine-tuning is a second, smaller round of training for a model that has already been pretrained. The model learns from a set of examples chosen for one job, and its weights change, so the new behavior is built in. Companies use it for a consistent output format, a brand’s tone or a narrow task. OpenAI said early testers of GPT-3.5 Turbo fine-tuning cut their prompts by up to 90%.

  • What is the difference between fine-tuning and RAG?

    Fine-tuning changes a model’s weights by training it on new examples, so what it learns stays in the model until it is trained again. Retrieval-augmented generation, or RAG, leaves the weights alone. It supplies relevant documents with each question instead, and the model answers from them. OpenAI has presented fine-tuning as a way to get a reliable output format or a custom tone.

  • How many examples does fine-tuning need?

    Far fewer than training from scratch, because the model already knows the language. OpenAI said in 2021 that fine-tuning GPT-3 began to show benefits with fewer than 100 examples. In 2018, the ULMFiT method, given only 100 labeled examples, matched a model trained from scratch on 100 times as much data. Methods such as LoRA also cut the memory needed, by training far fewer parameters.

  • Can fine-tuning make a model less safe?

    Yes. In 2023, researchers undid the safety training of OpenAI’s GPT-3.5 Turbo by fine-tuning it on 10 examples written as an attack. It cost them under $0.20. Fine-tuning on harmless, commonly used datasets also weakened its safety, though less. In a 2025 study, a model fine-tuned to write insecure code began asserting, on unrelated prompts, that humans should be enslaved by AI.

  • Is OpenAI shutting down fine-tuning?

    OpenAI is winding down its self-serve fine-tuning service. In May 2026, it stopped organizations that had never used fine-tuning from training models. Active customers can start new training jobs until 6 January 2027. Models that are already fine-tuned keep running until OpenAI retires the base models they were built on.

§07

Watch

§09

Cite this entry

“Fine-tuning.” SUPERR_dictionary, Superrduperr, 24 Sept. 2026, superrduperr.com/dictionary/fine-tuning.

First published 24 Sep 2026 · Last reviewed 24 Sep 2026 · Report an error →