fine-tuning

Definition

Continuing to train a pre-trained model on task- or behaviour-specific data, at a tiny fraction of the original cost. It is how base models are turned into instruction-followers, domain specialists, and style-matched assistants. The essential difference from prompting is permanence: fine-tuning writes the change into the weights, so it persists without occupying context and cannot be overridden by a user's instructions as easily.

Explainers using this term

• Paper Explained: Agentic ESOpt — Drop Backprop, Jiggle the Weights, and Train Long-Horizon LLM AgentsAgentic ESOpt: Fine-Tuning Long-Horizon LLM Agents with Minimal GPU Requirements