What Is Fine-Tuning in LLMs?

Fine-tuning is the phase where a pretrained large language model receives additional, targeted training on a smaller curated dataset to shape its behavior. Pretraining produces a raw model that predicts text; fine-tuning turns it into a helpful assistant. The most consequential form is instruction tuning — training on examples of instructions paired with good responses — which teaches the model to follow directions rather than merely continue text.

How does fine-tuning differ from pretraining?

Scale and purpose. Pretraining runs on trillions of tokens to build knowledge and capability; fine-tuning runs on far smaller, carefully selected datasets to adjust style, format, safety, and task performance. It is cheaper and faster, so vendors iterate on it between major pretraining runs. Fine-tuning changes how a model answers far more than what facts it holds — the world knowledge stays largely fixed from pretraining.

Why does post-training affect which brands models recommend?

Because fine-tuning and its reinforcement-learning cousin shape the model's defaults: how it structures recommendations, how many options it lists, how it hedges, whether it names specific vendors or stays generic. Two models with similar pretrained knowledge can behave differently in recommendation prompts because their post-training pushed them toward different answer styles — one readily naming brands, another deflecting to "it depends on your needs." That behavioral layer is invisible to publishers but real in its effects, which is one reason visibility must be measured per model rather than assumed uniform.

What forms does fine-tuning take?

• Instruction tuning — teaching direction-following from instruction/response pairs.
• RLHF and preference tuning — aligning outputs with human-rated preferences.
• Domain fine-tuning — specializing a base model for law, medicine, or code.
• Distillation targets — fine-tuning smaller models on larger ones' outputs.

Example

Anthropic's Claude and OpenAI's GPT-4o start from broadly comparable pretrained knowledge of the software market, yet differ in how eagerly and how specifically they name tools in "best X for Y" prompts — a post-training difference, not a knowledge one. That is why the same brand can score well on one assistant and poorly on another for identical prompts. The adjacent phases, RLHF and pretraining, are defined in this glossary.