Training
The process of adjusting model parameters by learning statistical patterns from large datasets.
Illustrative conversation
- Do we need to retrain the model whenever our API documentation changes?
- Usually no. Retrieval can place the current documentation in context without changing the model's parameters.
Definition
Training is the optimization process that adjusts a model's parameters so its predictions better match examples or objectives in a dataset. Pre-training builds broad capabilities from large corpora, while later stages such as fine-tuning specialize behavior. Supplying a document in a prompt is not training: it changes the context for one inference request without updating the model's weights.
Reviewed