Fine-tuning is the heavier-duty cousin of prompt configuration: instead of repeatedly asking a general model to behave differently, continue training a pretrained model on task- or domain-specific examples. The aim is specialization—more accurate, contextually relevant responses for a particular job. It is useful when an off-the-shelf model is not a good fit, but it is not a tiny settings tweak wearing a lab coat. section 1.7 section 9.8 section 9.4
Start with a pretrained language model, which has already learned broad language patterns, grammar, semantic representations, and knowledge across many topics. Then supply a task-specific dataset: examples or prompts paired with the outputs the model should produce. For sentiment analysis, that could mean sentences labeled positive or negative; for professional specialization, the data would lean heavily toward the relevant legal or medical content. section 1.7 section 1.7 section 9.5
A typical fine-tuning workflow has a few moving parts:
- Prepare labeled task data or prompt-and-response examples that represent the behavior you want. section 1.7
- Adapt or extend the model architecture when the task requires it, such as by changing task-specific, attention, or output layers. section 1.7
- Train on the task data, updating parameters with gradient-based optimization such as SGD or Adam to reduce the gap between predictions and desired outputs. section 1.7
- Repeat across multiple epochs and refine the model rather than expecting one pass to perform miracles. Computers, regrettably, remain immune to motivational speeches. section 1.7
Specializing for a professional domain
A model intended for lawyers or medical professionals needs a dataset rich in legal or medical material. That customization comes from the data used for continued training, not merely from choosing a base model. The practical catch is scale: fine-tuning requires access to your own datasets and significant hardware resources. section 9.5 section 9.8 section 9.4
Fine-tuning can sit beside prompt creation, but it reaches beyond ordinary prompt controls into model-level training and potentially architecture changes. It can also demand a GPU: a 24 GB Nvidia 3090 is described as sufficient for efficient LoRA fine-tuning in many LLM operations, while larger work may require multiple cards. Hardware, datasets, and rapidly changing implementation details make this a decision to investigate carefully rather than a casual afternoon configuration exercise. section 1.7 section 9.4 section 9.5
