Fine-Tuning
Fine-tuning further trains a pretrained model on a smaller, specific dataset to adapt its behavior for a particular task.
Prerequisites
What Is Fine-Tuning?
A foundation model is trained once, on broad general data, to be reasonably good at many things. Fine-tuning takes that already-trained model and continues training it on a smaller, more specific dataset — adjusting its weights so it performs better on a narrower task or adopts a particular style, format, or tone.
Foundation Model
Already TrainedTrained once on broad, general data.
Task-Specific Dataset
Narrow FocusSmaller and more specific than the original training data.
Further Training
Adjusts WeightsContinues training, changing the model's weights.
Fine-Tuned Model
SpecializedBetter at the narrower task, style, or tone.
Key Idea
Fine-tuning changes the model's weights. It's a different tool from prompting, which only changes what you send to an unchanged model.
Fine-Tuning vs. Prompting and RAG
Fine-tuning is one of several ways to adapt model behavior, and it's often not the first one to reach for:
No retraining required
Fast to change or iterate
Good for adding knowledge or instructions at request time
Limited by context window
Requires a training dataset and a training run
Slower to iterate — each change needs retraining
Good for teaching consistent style, format, or narrow behavior
Behavior is baked into the model itself
In practice, teams usually try prompting and RAG first, because they're faster to iterate on and don't require a training pipeline. Fine-tuning tends to be reserved for cases where the model consistently needs to behave a certain way that prompting alone can't reliably achieve — a particular output format, domain-specific tone, or a narrow, repeated task.
A Real-World Example
A support team wants every response to follow a strict internal format and tone. Prompting can request this, but a general-purpose model may still drift from the format on edge cases. Fine-tuning the model on hundreds of examples of correctly formatted responses can make that format far more consistent, because the desired behavior becomes part of the model's learned weights rather than something re-requested on every call.
Common Mistakes
Reaching for fine-tuning to add knowledge
RAG is usually a better fit for giving a model access to specific facts or documents — fine-tuning is about adjusting behavior and style, not injecting a knowledge base.
Fine-tuning before trying prompting or RAG
Fine-tuning is slower and more expensive to iterate on. Most teams should exhaust simpler approaches first.
Assuming fine-tuning fixes hallucination
A fine-tuned model can still generate confident, incorrect output — fine-tuning changes style and task performance, not whether the model knows something is true.
Using too little or low-quality training data
A small or inconsistent dataset can make the fine-tuned model unreliable or teach it the wrong pattern.
Forgetting the model still needs to be evaluated
A fine-tuned model should be measured against real examples, not assumed to have improved just because training completed.
Interview Question
What is fine-tuning, and when would you use it instead of prompting or RAG?
Fine-tuning continues training an already-pretrained model on a smaller, task-specific dataset, adjusting its weights so it behaves differently by default — as opposed to prompting or RAG, which change what you send to an unchanged model. You'd reach for fine-tuning when you need consistent style, format, or narrow task behavior that prompting alone can't reliably produce, and you have enough quality training examples to teach that pattern. It's usually not the first tool to try, since it's slower and more expensive to iterate on than prompting or RAG, and it's not a good fit for injecting new factual knowledge — RAG handles that better.
What an interviewer may ask next
- Why would you try prompting or RAG before fine-tuning?
- Why is RAG usually a better fit than fine-tuning for giving a model access to specific documents?
- Does fine-tuning reduce hallucination? Why or why not?
Explain It in 30 Seconds
Fine-tuning takes an already-trained foundation model and continues training it on a smaller, specific dataset, adjusting its weights to change its default behavior — style, format, or a narrow task. It's different from prompting or RAG, which change what you send to an unchanged model. Most teams try prompting and RAG first, since they're faster to iterate on, and reach for fine-tuning when they need consistent behavior that prompting can't reliably achieve.