Fine-tuning is the process of taking a pre-trained machine learning model and continuing its training on a smaller, task-specific dataset so that it performs better on a narrower domain or use case. Rather than building a model from scratch, fine-tuning leverages the broad knowledge already encoded in a foundation model and adapts it to a particular context, such as customer support conversations, legal document summarization, or medical diagnosis assistance.
To understand why fine-tuning matters, it helps to understand how large language models (LLMs) and other AI models are built. A foundation model is trained on enormous volumes of general data, a process that requires massive computational resources and can take weeks or months. Fine-tuning allows organizations to take this pre-built foundation and specialize it using their own proprietary data, at a fraction of the original cost. The resulting model retains its general capabilities while gaining a sharper understanding of the specific vocabulary, tone, or reasoning patterns present in the new training set.
Fine-tuning is often compared to two other techniques for customizing model behavior: prompt engineering and retrieval-augmented generation (RAG). Prompt engineering shapes model output by crafting carefully worded instructions at inference time, without changing the model's underlying weights. RAG supplements a model's responses by retrieving relevant documents from an external knowledge base at query time. Fine-tuning, by contrast, actually modifies the model's parameters during an additional training phase. This makes it more suitable for instilling consistent behavior, style, or specialized knowledge that cannot be reliably conveyed through prompts alone.
The practical trade-offs between these approaches are significant. Fine-tuning requires labeled training data, computational resources for the training run, and ongoing maintenance if the target domain evolves. RAG, on the other hand, can be updated simply by refreshing the underlying document store. For many applications, the two techniques are used together: a fine-tuned model handles domain-specific language and reasoning, while RAG provides access to current or proprietary information.
A closely related concept is reinforcement learning from human feedback (RLHF), which is itself a form of fine-tuning. In RLHF, human raters evaluate model outputs, and those preferences are used to guide further training, steering the model toward responses that are more helpful, accurate, or aligned with intended behavior. This technique was central to the development of instruction-following models such as GPT-3.5 and later versions.
Fine-tuning is also sometimes referred to as model adaptation or transfer learning, though transfer learning is the broader concept from which fine-tuning descends. In transfer learning, knowledge gained from one task is applied to a different but related task. Fine-tuning is the most common practical implementation of that principle in modern AI development.