Fine-Tuning
The process of further training a pre-trained model on a smaller, task-specific dataset to adapt it for a particular use case or domain.
When to Fine-Tune
When prompting alone doesn't achieve desired quality, when you need consistent behavior for a specific task, when domain-specific knowledge is required, or when you need to reduce inference costs by using a smaller specialized model.
Methods
Full fine-tuning: Update all parameters (expensive, risk of forgetting). LoRA/QLoRA: Only update small adapter matrices (efficient, preserves base knowledge). Prompt tuning: Only learn soft prompt tokens.
Best Practices
Start with a strong base model. Use high-quality, diverse training data. Monitor for overfitting with validation metrics. Use parameter-efficient methods when possible. Evaluate on held-out data that reflects real use.