The vocabulary
Plain-language definitions, each linked to where the idea lives in the field.
Running a trained model to get predictions.
Reinforcement learning from human feedback — tuning a model toward preferred answers.
Further training a pretrained model on a smaller, specific dataset.
The neural network architecture behind most modern language models.
When a model states something fluent but false.
How much text a model can consider at once.
A chunk of text (often a word-piece) that a model reads and predicts.
A list of numbers that represents meaning, so similar things sit close together.
Shrinking a model by storing its weights at lower precision.
Generating data by gradually removing noise, step by step.
Making a model's behavior match human intent and values.
A mechanism that lets a model weigh which parts of the input matter most for each output.
