Fine tuning refers to the subsequent optimization of a pre-trained AI model for a specific task or vocabulary.
For the original training of these models, large amounts of data from the Internet were first processed to gain a general understanding. Fine tuning then adapts this general knowledge to a specific use case.
For example, an insurance company may want to use an AI-based chatbot to communicate with customers. The existing language model is retrained based on insurance documents, customer chats, and technical terms used by the company. This gives the chatbot the domain knowledge it needs to accurately answer insurance questions.
Unlike the initial training, the fine-tuning typically only involves optimizing the weights between the model’s neurons. The architecture itself remains the same. The advantage is that performance can be significantly improved with comparatively little additional training data.
The difference to RAG (Retrieval Augmented Generation) is that fine-tuning involves a permanent adaptation of the language model. Instead, RAG queries additional information sources in real time for each query. There is no new training.