Training

Training is the process of learning or updating a model's parameters (weights) from data by optimizing an objective (loss), typically using gradient-based methods. It is primarily an ML engineering concern; AI engineering consumes trained models rather than producing them.

Details

The distinction from inference (which uses weights as-is) is fundamental to AI engineering: AI engineers work at inference time with models whose behavior was shaped during training, and understanding the training pipeline helps explain why models behave as they do.

Training encompasses pretraining (learning general representations from broad data), post-training phases such as fine-tuning and reinforcement learning, and alignment work. Training data can include synthetic data generated by other models. Capabilities like tool calling and instruction following are established or refined during training.

Synonyms

model training