Fine-Tuning
Also: Model fine-tuning, Domain adaptation, Supervised fine-tuning (SFT)
Fine-tuning adapts a pre-trained model to a specific task or domain by continuing training on a smaller, targeted dataset, improving accuracy and style for that use case.
What It Is
Fine-tuning is the process of taking a model that has already been trained on a large, general dataset (a pre-trained or foundation model) and continuing its training on a smaller, focused dataset. The goal is to specialize the model so it performs better on a particular task, domain, or style without training a new model from scratch.
Because the base model already encodes broad knowledge (language patterns, visual features, or general reasoning), fine-tuning only needs to adjust the model to the new objective. This makes it far cheaper and faster than full training.
Why it matters
- Lower cost and data needs: You reuse the expensive pre-training, so a few hundred to a few thousand labeled examples can be enough.
- Better task performance: A fine-tuned model usually beats a generic model on niche tasks (legal summaries, medical coding, brand voice).
- Consistency: It can lock in a tone, format, or domain vocabulary that prompting alone struggles to enforce reliably.
How it is used in practice
1. Collect data: Assemble high-quality input/output pairs representative of the target task.
2. Choose a method:
- Full fine-tuning: update all model weights (most flexible, most expensive).
- Parameter-efficient fine-tuning (PEFT): update only small adapter layers, for example LoRA, leaving most weights frozen.
3. Train: Run additional training epochs with a low learning rate to avoid erasing prior knowledge (a risk called catastrophic forgetting).
4. Evaluate and deploy: Validate on held-out examples, then serve the adapted model.
Concrete Example
A support team has a general language model that answers questions but does not match their product. They prepare 1,500 past tickets paired with approved agent responses. After fine-tuning, the model replies using the company's terminology, escalation rules, and tone. Accuracy on internal evaluations rises from 68% to 91%, and responses need less editing.
When not to use it
If your need is occasional or knowledge changes often, prompt engineering or retrieval augmented generation (RAG) may be cheaper and easier to update than retraining.