Fine-Tuning
Fine-tuning is the process of further training an existing AI model on additional examples so its behaviour becomes better suited to a particular task or pattern.
It modifies the model rather than simply giving it more information during one request.
What It Can Be Used For
Fine-tuning may help with tasks such as:
- Consistent classification
- Specific output formats
- Specialized terminology
- Repeated behavioural patterns
- Particular styles or conventions
- Tool-selection behaviour
The suitability depends on the model and provider.
Fine-Tuning vs. Prompting
Prompting provides instructions during the interaction.
Fine-tuning adjusts model behaviour through additional training.
If a prompt reliably produces the required behaviour, fine-tuning may add unnecessary complexity.
Fine-Tuning vs. RAG
Retrieval-Augmented Generation supplies external knowledge at runtime.
Fine-tuning is usually better understood as a way of adapting behaviour.
A common mistake is attempting to fine-tune a model simply to:
teach it our current documentation.
If the underlying information changes regularly, retrieval is often easier to update.
Why It Matters
Fine-tuning creates another model dependency that needs to be:
- Evaluated
- Versioned
- Maintained
- Recreated when base models change
- Documented
It is therefore not necessarily the first optimization step.
Current OpenAI guidance separates retrieval techniques for improving knowledge and accuracy from fine-tuning techniques intended to improve behaviour or consistency. OpenAI model-selection guidance
Common Problems
Common problems include:
- Fine-tuning before improving the Prompt
- Using poor training examples
- Expecting fine-tuning to create a reliable factual database
- No held-out evaluation set
- Training data not representative of production
- No documentation of which base model was used
- No migration plan when the model changes
How to Improve
Start with AI Evaluation.
Identify the repeated failure pattern.
Then determine whether the problem is actually:
- Prompting
- Retrieval
- Source data
- Tool logic
- Model selection
- Or model behaviour appropriate for fine-tuning
Only introduce the additional training layer when it solves a clearly observed problem.