An adapter is trained on examples, then loaded later to influence generation.
Training changes learned parameters. Pretraining establishes a base model; fine-tuning adapts an existing one using further data. Hugging Face explains fine-tuning. The required data, objective and computation depend on the model and task.
Do not confuse training with supplying a prompt. The model can use instructions for the current answer without saving them into its weights. Adding a document to a RAG search index changes retrieval material rather than automatically retraining the language model.
First identify the problem. Missing facts may be handled by reliable context. An output format may be addressed through instructions and validation. Consistent adaptation across many examples may justify a separate experiment. Keep training data, a held-out evaluation set and the saved version distinct so you can assess whether the change helped.
Sources
Primary references checked 11 October 2026. This explanation is not a benchmark of a particular model or computer.