Fine-tuning adapts a pre-trained language model to perform better on specific tasks. While many use cases can be solved with prompt engineering alone, fine-tuning offers significant advantages for specialized applications.

When to Fine-Tune

Good reasons to fine-tune:

  • Your domain has specialized vocabulary or formats
  • You need consistent output formatting
  • You want to reduce latency and cost compared to prompt-based approaches
  • You're building a product that requires specialized behavior

When NOT to fine-tune:

  • Your use case can be solved with good prompts
  • You don't have sufficient high-quality training data
  • You need to change behaviors frequently

Methods

Full fine-tuning: Update all model parameters. Expensive but most powerful.

LoRA (Low-Rank Adaptation): Train small adapter layers while freezing the base model. Much cheaper, often equally effective.

QLoRA: Quantized LoRA that can run on consumer GPUs.

Data Preparation

Quality matters more than quantity. A few hundred high-quality examples often outperform thousands of noisy ones. Ensure your data:

  • Represents real use cases
  • Has consistent formatting
  • Is free of errors and biases

Evaluation

Always evaluate your fine-tuned model against a held-out test set. Common metrics include accuracy, F1 score, and human evaluation.

Production Considerations

  • Versioning: Track which base model and data were used
  • Monitoring: Watch for drift in performance
  • Cost: Fine-tuned models are cheaper per-token than large general models

Learning More

Our Build with LLMs programme covers fine-tuning with hands-on projects using modern tools and techniques.