Glossary
What is Fine-tuning?
Subjecting a pre-trained model to additional training with a smaller, specialized dataset for a specific domain or task.
Fine-tuning is adapting a ready model to a particular tone, terminology or task, instead of training a general-purpose LLM from scratch. The model parameters are updated with the organization’s examples.
This method is effective for changing the model’s behavior and format; it is generally not suitable for conveying frequently changing information. For current documents, the RAG approach is in most cases more practical and does not require retraining for data updates. The quality of the training data and its freedom from personal data are critically important.
Fine-tuning requires GPU resources, but it is much lighter than training from scratch. The resulting model can be run in-house for inference. When deciding, trying prompt design and RAG first is a common order.