33 articles
Browse guides
Parameters are the learnable values inside a model — mostly weights and biases. Training is the process of nudging them until the model’s outputs stop being wrong. Headline counts like ‘70B parameters’ describe capacity, not quality: architecture, data, and post-training often matter more than raw size. In practice: A 70B model has roughly 70 billion adjustable numbers; a bigger number does not automatically mean better answers.
Precision asks: of everything the model flagged, how much was right? Recall asks: of everything it should have flagged, how much did it catch? They trade off against each other, and which one you optimise is a product decision, not a technical one. Accuracy hides both. In practice: A fraud model catching 60% of fraud (recall) with 90% of its alerts genuine (precision) may beat a 99.9%-accurate one that catches nothing. ...
Regression models output a continuous value — a price, a temperature, a days-to-churn estimate. The contrast with classification is what you are predicting, not how the model works internally. In practice: Estimating that a flat will sell for €340,000.
Regularisation trades a little training accuracy for better generalisation. Common forms are penalising large weights, dropping random units during training, and stopping early. It is the main lever against overfitting when you cannot get more data. In practice: Dropout randomly switches off units each step so the network cannot rely on any single one.
A reinforcement learning agent acts in an environment, observes what happens, and receives a reward signal. Over many episodes it learns a policy that maximises expected reward. RL is behind game-playing systems and, in modified form, behind the alignment step in modern chat models. In practice: An agent learns to play a game with no rules explained — only a score that goes up or down.
Self-supervised learning invents a prediction task out of unlabelled data, most famously ‘predict the next token’. Because the label is just the next word, the entire internet becomes training data with no human annotation. This is why LLMs could scale in a way supervised models never could. In practice: Hide the last word of a sentence, ask the model to guess it, repeat trillions of times.
In supervised learning each training example is a pair: an input and its label. The model predicts, compares against the label, and corrects itself. It is the most common and most reliable setup, and its main constraint is that labels are expensive to produce. In practice: 10,000 emails each tagged spam or not-spam, used to train a filter.
Test data is the honest scorecard. It must stay untouched during development, because any decision made on the basis of test results turns it into validation data. When test examples leak into training — contamination — reported scores become meaningless. In practice: A model scoring 95% on a benchmark whose answers were in its training data has learned nothing.
Training data is the corpus used to fit a model’s parameters. Its coverage sets the model’s ceiling: gaps, errors, and skew in the data reappear as gaps, errors, and bias in the outputs. Under the EU AI Act, training data has a formal definition and, for high-risk systems, governance obligations attached. In practice: A hiring model trained mostly on past hires from one demographic will reproduce that pattern.
Underfitting is the opposite failure to overfitting: the model has not learned enough. Symptoms are poor scores everywhere, and the fixes are more capacity, better features, longer training, or a less aggressive regulariser. In practice: Fitting a straight line to data that clearly curves.