Fine-Tuning

Fine-tuning takes a general pretrained model and teaches it something more specific: following instructions, writing in a company’s voice, classifying insurance claims, or speaking a medical specialty’s jargon.

Common types

• Supervised fine-tuning (SFT): training on example inputs paired with ideal outputs, such as thousands of well-written assistant replies.
• Preference tuning: training on which of two answers people prefer, as in RLHF.
• Parameter-efficient fine-tuning (e.g., LoRA): adjusting a small add-on set of weights instead of the whole model, which is much cheaper.

Fine-tuning vs alternatives

Need	Often best approach
Model needs up-to-date or private facts	Retrieval (RAG)
Consistent format, tone or narrow skill	Fine-tuning
Quick behaviour change	Prompt engineering

Who can fine-tune

With open-weights models, anyone with suitable hardware can fine-tune. Some closed-model providers offer fine-tuning through their APIs. Thinking Machines Lab’s Tinker service (2025) was built specifically to make fine-tuning open models easier.

A safety caveat

Fine-tuning can also remove safety training from open models. That is one reason the release of very capable open weights is debated.