Training
Training is the process that turns an untrained model into a working AI: you feed it thousands or millions of examples, compare its guesses with the correct answers, and nudge its internal settings until the guesses get good.
No jargon, no math
In plain words, what is it?
Training works like flashcards that grade themselves. You start with a model full of random settings, its parameters. Each example goes in, the model makes a guess, and the training loop measures how wrong it is, the loss. Then it adjusts every parameter a tiny amount, in the direction that makes the next guess less wrong. That adjustment is gradient descent, computed on GPUs thousands of times per second. Repeat over millions of examples and several passes, called epochs, and the guesses stop being guesses: they become predictions. The boundary matters: train too long on too little data and the model memorises the examples instead of learning the pattern, overfitting. A locked-away test set is how you catch that before it ships.
- 01
Feed examples. Clean examples enter as tokens, one batch at a time.
- 02
Guess. The model predicts the answer with its current parameters.
- 03
Score. The loss function measures how far the guess is from correct.
- 04
Adjust. Gradient descent nudges every parameter a little toward less error.
- 05
Repeat and check. Batches loop until the error is low, and a locked test set confirms the model learned the pattern: then it is ready.
Terms that live nearby.
Dataset
A dataset is the organised collection of examples an AI learns from or is judged on: thousands or millions of rows of text, images or sensor readings, cleaned, labelled and split into training, validation and test parts.
ReadParameter
A parameter is one of the internal numbers a trained model stores: billions of them, each tuned during training, and together they encode everything the model learned.
ReadOverfitting
Overfitting is when a model memorises its training data instead of learning the underlying pattern: near-perfect on examples it has seen, unreliable on anything new.
Read§ 04
Want this running in your stack?
We design and ship AI systems with permissions, logs and human approval. It starts with a 30-minute scoping call.