Model mechanics and performance
What is inside the model and how its behaviour is measured and goes wrong. Variables live in the data; parameters and weights live in the model - and overfitting (memorising) vs underfitting (too simple) is a classic confusion pair.
This cluster covers what is inside the model and how its behaviour is measured - and how it goes wrong. Start with the parameter family, because the micro-distinction is a planted exam point.
- Parameters → the internal values a model learns during training, counted in billions for LLMs and a rough proxy for scale.
- Weights → the learned numeric strengths on connections deciding each input's influence - the core kind of parameter.
- Inference → the production phase where the trained model applies what it learned to new inputs.
- Generalization → performing well on new, unseen data, not just the training set.
- Variance → how sensitive outputs are to fluctuations in the training data; high variance is overfitting territory.
- Entropy → a measure of uncertainty or randomness, used, e.g., to choose decision-tree splits.
- Accuracy → the share of outputs that are correct against ground truth.
- Hallucinations → GenAI output that contradicts the source or is factually wrong while presented as fact.
- Counterfactual → an explanation showing the minimal input change that would flip the output ("had income been higher, the loan would be approved").
Overfitting → the model memorises the training data, noise included - brilliant in training, poor on new data. Underfitting → too simple to capture the pattern, so it fails on training AND new data. Overfit fails on the new; underfit fails everywhere.
Memory hook for the parameter family: variables live in the data; parameters and weights live in the model - weights being the learned connection strengths that make up most of the parameters.