Fine-tuning is the practice of taking a model that has already been trained, usually on a large, general-purpose dataset during Pretraining, and continuing its training on a smaller, more specific dataset so that it adapts to a particular task, domain, or style. The technique rests on the premise of Transfer learning: a model that has already learned broad statistical structure, whether visual features from images or linguistic patterns from text, needs comparatively little additional data and compute to specialize, compared to training a model of similar quality from scratch.
History
Fine-tuning became a standard practice in computer vision after the 2012 ImageNet breakthrough, when researchers found that convolutional networks trained on ImageNet transferred well to smaller, unrelated image classification tasks with only light retraining of the final layers. In natural language processing, the technique took on new significance with BERT in 2018, whose "pretrain then fine-tune" paradigm, adapting a single pretrained encoder to many downstream tasks such as sentiment analysis or question answering, became the dominant workflow for several years. With the rise of large large language models, fine-tuning shifted from adapting models for narrow classification tasks toward instruction fine-tuning, where a base model trained purely to predict the next token is adjusted on curated examples of instructions and desired responses to make it behave as a helpful assistant, a step used in training systems such as ChatGPT.
Methods
Full fine-tuning updates all of a model's parameters and, for the largest models, requires substantial compute and memory, since it involves storing gradients and optimizer states for every weight. This cost drove the development of parameter-efficient fine-tuning methods that freeze most of the original model and train only a small number of additional parameters. The most widely adopted of these is LoRA, which inserts small low-rank matrices into a model's layers and trains only those, cutting memory requirements dramatically while achieving results close to full fine-tuning on many tasks. Other approaches include adapter layers, prompt tuning, and prefix tuning, each trading off a different amount of flexibility against efficiency.
Supervised fine-tuning (SFT) trains a model on labeled input-output pairs and is the standard first step in adapting a base LLM into an instruction-following assistant. This is often followed by preference-based methods such as RLHF or Direct Preference Optimization that further refine the model's behavior using human judgments about which of two outputs is better, rather than a single correct label.
Fine-tuning versus other approaches
Fine-tuning is frequently compared with Retrieval-augmented generation as a way to make a model produce more accurate or specialized outputs. Fine-tuning changes the model's internal weights and is well suited to teaching a consistent style, format, or specialized skill; retrieval instead supplies relevant facts at inference time without altering the model, making it easier to update as source information changes and less prone to overwriting the model's general capabilities. In practice, the two are often combined, with a fine-tuned model calling on retrieved documents for grounding.
Risks
Fine-tuning on a narrow dataset can cause catastrophic forgetting, where the model loses general capabilities it had before adaptation, and Overfitting to a small fine-tuning set is a persistent risk, particularly when the number of examples is small relative to the model's capacity. Fine-tuning on unfiltered or poorly curated data has also been shown to weaken a model's safety training, a concern that has driven providers to restrict or monitor fine-tuning access to their most capable models.