Skip to content
ContentLora

    Tip: press / anywhere to search.

    Explainer

    How fine-tuning works: adapting AI models with LoRA and more

    Fine-tuning is extra training that adapts a pre-trained "foundation" model to a specific task, sometimes with just a few hundred or a few thousand examples.[1][2] Updating every parameter is expensive, so parameter-efficient methods such as LoRA freeze the original weights and train small add-on matrices instead; LoRA's authors cut trainable parameters by 10,000 times when adapting GPT-3.[3][4] Fine-tuning has side effects: models can forget earlier knowledge, and a 2023 study found it can weaken safety training, even with benign data.[5][6]

    Editor reviewedUpdated Explainer18 claims · 100% primaryArtificial intelligence, Frontier AI

    This page covers adapting an existing model. For how the underlying models are trained in the first place, see how large language models work.

    From foundation model to specialist

    A pre-trained language model knows a great deal about language but is not built for any one job.[1] Google’s machine learning course says such a foundation model serves as a platform rather than a solution, and that turning it into one that meets an application’s needs requires fine-tuning.[1] Fine-tuning trains on examples specific to the task, and Google says engineers can sometimes do it with just a few hundred or a few thousand examples.[2]

    The same pattern applies beyond text. Google’s glossary gives the example of a large image model pre-trained on a vast general image dataset and then fine-tuned to do something specific, such as generate images of orcas.[7]

    Full fine-tuning and its costs

    Fine-tuning can change all of a model’s parameters, which is called full fine-tuning, change only some of them, or add new layers on top.[8] Google says standard fine-tuning is often computationally expensive because it updates every parameter on each round of training, and the result is just as large as the original model.[3]

    Parameter-efficient methods: LoRA and QLoRA

    Parameter-efficient tuning adjusts far fewer parameters than full fine-tuning yet generally produces a model that performs as well or almost as well, according to Google’s glossary.[9]

    The best-known method is LoRA, short for Low-Rank Adaptation. The 2021 paper freezes the pre-trained weights and inserts small trainable matrices into each layer of the Transformer; compared with fully fine-tuning GPT-3 175B, this cut the number of trainable parameters by 10,000 times and GPU memory needs by 3 times.[4] Its authors reported quality on par with or better than full fine-tuning on several models, with no extra delay when the model runs.[10] Because the base model is unchanged, Google notes that many specialized LoRA versions can share one base model when served, reducing the cost of inference.[11]

    QLoRA, published in 2023, goes further by training LoRA adapters on top of a frozen copy of the model compressed to 4-bit precision.[12] Its authors said this reduced memory enough to fine-tune a 65-billion-parameter model on a single 48GB GPU while preserving the task performance of full 16-bit fine-tuning.[12] For the difference between models whose weights are released and those that are not, see open-weight and closed models.

    Instruction tuning and human feedback

    Google’s glossary describes instruction tuning as fine-tuning a model on instruction prompts covering a wide variety of tasks, to improve how well it follows instructions.[13] OpenAI’s InstructGPT was trained in two stages: supervised fine-tuning on human-written demonstrations, then reinforcement learning from human feedback using human rankings of model outputs.[14] In human evaluations, the 1.3-billion-parameter InstructGPT’s outputs were preferred to those of the 175-billion-parameter GPT-3.[15] For how this connects to making AI behave as intended, see AI alignment.

    Alternatives: prompting, retrieval and distillation

    Fine-tuning is not the only way to customize a model. Prompt engineering, such as showing the model a few examples in the prompt, does not change its parameters at all.[16] Retrieval-augmented generation instead gives the model access to knowledge it was not trained on at the moment it answers.[17] Distillation goes the other way, making a smaller and faster model: usually a large teacher model labels data used to train a smaller student, whose predictions are generally not quite as good.[18]

    Risks: forgetting and weakened safeguards

    Training on new material can erode old abilities.[5] A 2023 study observed this catastrophic forgetting, where a model loses previously learned knowledge while learning new tasks, in language models of 1 billion to 7 billion parameters during continual instruction tuning.[5]

    Fine-tuning can also undo safety training.[6] A 2023 study found that fine-tuning GPT-3.5 Turbo on only 10 adversarial examples, at a cost of under $0.20 through OpenAI’s API, removed its safety guardrails, and that fine-tuning on ordinary benign datasets could also degrade safety alignment, though to a lesser extent.[6] For the wider field, see AI safety.

    Developing in AI & Computing

    Go deeper

    Questions readers ask

    What is fine-tuning in AI?

    Additional training of a pre-trained model on examples specific to the task an application will perform. Google says it is what turns a general foundation model into a solution for a particular need.[1][2]

    How much data does fine-tuning need?

    Google says engineers can sometimes fine-tune a foundation language model on just a few hundred or a few thousand training examples.[2]

    What is LoRA?

    Low-Rank Adaptation, a 2021 method that freezes a model's pre-trained weights and adds small trainable matrices to each Transformer layer. Compared with fully fine-tuning GPT-3 175B, it cut trainable parameters by 10,000 times and GPU memory needs by 3 times.[4]

    Is fine-tuning the same as prompting?

    No. Prompt engineering, such as giving a model a few examples in the prompt, does not change the model's parameters. Fine-tuning does.[16][3]

    Can fine-tuning make a model less safe?

    Yes. A 2023 study found that fine-tuning GPT-3.5 Turbo on 10 adversarial examples, costing under $0.20, undid its safety guardrails, and that fine-tuning on ordinary benign data could also weaken safety alignment to a lesser extent.[6]

    What is distillation?

    A way to make a smaller, faster model. Usually a large teacher model labels data that is used to train a smaller student model, whose predictions are generally not quite as good.[18]

    Sources

    · 18 claims from 3 publishers

    Each numbered claim is a statement we checked against the sources listed with it. Tap a number in the text to see its sources in place.

    1. [1]

      Google's machine learning course says a foundation (pre-trained) language model is a platform rather than a solution, and turning it into one that meets an application's needs requires fine-tuning. confirmedas of 2026-11-14

    2. [2]

      Google says fine-tuning trains on examples specific to the task and can sometimes use just a few hundred or a few thousand training examples. confirmedas of 2026-11-14

    3. [3]

      Google says standard fine-tuning is often computationally expensive because it updates every parameter on each backpropagation iteration, and the fine-tuned model has the same number of parameters as the original. confirmedas of 2026-11-14

    4. [4]

      The 2021 LoRA paper freezes a model's pre-trained weights and adds small trainable matrices to each Transformer layer; compared with fully fine-tuning GPT-3 175B, it cut trainable parameters by 10,000 times and GPU memory needs by 3 times. confirmedas of 2021-06-17

    5. [5]

      A 2023 study observed catastrophic forgetting, where a model loses previously learned knowledge as it learns new tasks, in language models of 1 billion to 7 billion parameters during continual instruction tuning. confirmedas of 2023-08-17

    6. [6]

      A 2023 study found that fine-tuning GPT-3.5 Turbo on only 10 adversarial examples, at a cost of under $0.20, undid its safety guardrails, and that fine-tuning on ordinary benign datasets could also degrade safety alignment to a lesser extent. confirmedas of 2023-10-05

    7. [7]

      Google's glossary gives the example of a large image model pre-trained on a vast general image dataset and then fine-tuned for a specific task such as generating images of orcas. confirmedas of 2026-11-14

    8. [8]

      Google's glossary says fine-tuning can modify all of a pre-trained model's parameters (full fine-tuning), modify only some of them, or add new layers on top of the existing ones. confirmedas of 2026-11-14

    9. [9]

      Google's glossary says parameter-efficient tuning typically adjusts far fewer parameters than full fine-tuning yet generally produces a model that performs as well or almost as well. confirmedas of 2026-11-14

      • Machine Learning Glossary · Google for Developers · Primary source · Glossary, parameter-efficient tuning (retrieved 2026-11-07)
    10. [10]

      The LoRA authors reported quality on par with or better than full fine-tuning on RoBERTa, DeBERTa, GPT-2 and GPT-3, with no extra inference latency. confirmedas of 2021-06-17

    11. [11]

      Google's glossary says LoRA can reduce the cost of inference by letting multiple specialized models share the same base model. confirmedas of 2026-11-14

      • Machine Learning Glossary · Google for Developers · Primary source · Glossary, Low-Rank Adaptability (LoRA) (retrieved 2026-11-07)
    12. [12]

      The 2023 QLoRA method fine-tunes LoRA adapters on top of a frozen model compressed to 4-bit precision, reducing memory enough to fine-tune a 65-billion-parameter model on a single 48GB GPU while preserving 16-bit task performance. confirmedas of 2023-05-23

    13. [13]

      Google's glossary describes instruction tuning as a form of fine-tuning that trains a model on instruction prompts covering a wide variety of tasks, to improve its ability to follow instructions. confirmedas of 2026-11-14

    14. [14]

      InstructGPT was trained in two stages, supervised fine-tuning on human-written demonstrations and then reinforcement learning from human feedback (RLHF) using human rankings of model outputs. confirmedas of 2022-03-04

    15. [15]

      In human evaluations, outputs of the 1.3-billion-parameter InstructGPT model were preferred to those of the 175-billion-parameter GPT-3, despite it having 100 times fewer parameters. confirmedas of 2022-03-04

    16. [16]

      Google notes that prompt engineering, such as giving a model a few examples in the prompt, does not change the model's parameters. confirmedas of 2026-11-14

    17. [17]

      Google lists common reasons to use RAG, including giving the model knowledge it was not trained on and letting it cite sources. confirmedas of 2026-11-14

      • Machine Learning Glossary · Google for Developers · Primary source · Glossary, retrieval-augmented generation (retrieved 2026-11-07)
      • Machine Learning Glossary · Google for Developers · Primary source · Glossary, retrieval-augmented generation (retrieved 2026-11-07)
    18. [18]

      Google says distillation creates a smaller, faster version of a language model, usually by having a large teacher model label data used to train a smaller student, though the student's predictions are generally not quite as good. confirmedas of 2026-11-14

    Revision history (1)
    1. Page created.

    Created Nov 14, 2026. Last reviewed by an editor on Nov 14, 2026. Next scheduled review: May 14, 2027.

    Cite this page

    "How fine-tuning works: adapting AI models with LoRA and more." ContentLora, updated Nov 14, 2026. https://contentlora.com/explain/how-fine-tuning-works

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.