Fine-tuning

Fine-tuning continues training a pre-trained model on a smaller domain-specific dataset to adapt its behavior, style, or capabilities for a particular use case.

What is fine-tuning?

Fine-tuning takes a pre-trained foundation model and continues training on a smaller, focused dataset to adapt it for a specific use case. The full-parameter variant (Supervised Fine-Tuning, SFT) is expensive but capable. Parameter-efficient variants (LoRA, QLoRA, adapter layers) tune a small fraction of parameters, are cheaper, and produce smaller deployable artifacts.

When to fine-tune vs RAG

Use fine-tuning when you need consistent style or specialized capability that doesn't fit in a prompt window. Use RAG when the underlying knowledge changes frequently or when you need to cite sources. Many production deployments use both. A naive shortcut — fine-tune on your knowledge base to "encode" the data — usually performs worse than RAG over the same data and is harder to update when the knowledge changes.

Procurement questions

If a vendor offers fine-tuning on customer data: who owns the resulting model weights, is the model isolated to the customer or pooled across customers, what happens to the fine-tuned model when the contract ends, are the training prompts and labels retained, and how is the customer informed when the base model is updated. Pooled fine-tuned models can leak signal across customers; isolated models are safer but more expensive to maintain.