Fine-Tuning

General definition

Fine-tuning starts from a foundation model that has already learned general language from a large corpus. A smaller, labelled dataset is then used to update the model’s weights so it performs better on a narrower task, such as clinical note summarisation, legal contract review, or customer support routing in a specific product.

Fine-tuning sits between two lighter alternatives: prompt engineering (steering behaviour through instructions without changing weights) and retrieval-augmented generation (grounding answers in retrieved documents). It is appropriate when the target domain has distinctive language patterns or when consistent output format matters more than access to up-to-date information.

• Requires labelled training examples, typically hundreds to tens of thousands depending on task complexity
• Updates model weights, so it demands GPU compute during training but no extra retrieval step at inference time
• Produces a separate model variant that must be versioned, evaluated, and monitored independently
• Lower-cost alternatives such as LoRA and QLoRA allow fine-tuning a fraction of a model’s parameters, reducing the GPU memory required

In the Ethora ecosystem

When you deploy a self-hosted LLM agent through the Ethora AI SDK, you control which model backs the agent. A fine-tuned model can be dropped in via a compatible API endpoint, giving your agent specialised behaviour without sending data to a third-party cloud.

For regulated verticals such as healthcare and finance, fine-tuning on a private dataset is often preferable to relying on a general model accessed over the public internet, because it keeps sensitive training data inside your own infrastructure.