Fine-tuning
The process of additionally training a pre-trained AI model on a specific dataset to optimize it for a particular task or domain. It is used to precisely adjust the required expertise or response style for a specific purpose while maintaining the general knowledge of the foundation model.
Detailed explanation
Why it matters for tool selection
Support for fine-tuning is crucial when you need to consistently apply a model's 'tone & manner', 'output format (JSON, code, etc.)', or 'specific business rules' rather than just retrieving knowledge. Particularly in security-conscious on-premise environments, a recent trend is fine-tuning small language models (SLMs) to achieve specialized expertise comparable to high-performance models.
What to check
- Do you have a high-quality, cleaned dataset (at least hundreds to thousands of examples) prepared?
- Is it a simple knowledge retrieval problem that can be solved with RAG (Retrieval-Augmented Generation) alone?
- Can you afford the GPU resources or API costs required for training?
- Will you use an open-source model where you can directly modify the model's weights?
Examples
An example is a financial institution fine-tuning a general GPT model on customer service logs to create a dedicated chatbot that accurately uses financial terminology and speaks in the company's distinct, polite tone when declining requests.
Commonly confused terms
RAG (Retrieval-Augmented Generation)
A technique that retrieves and references external knowledge to generate answers, without altering the model's underlying intelligence or personality.
Prompt Engineering
The skill of guiding a model through prompt design, adjusting the output based on the input alone without any additional training.