A training technique where parts of text are hidden and the model learns to predict what should fill those gaps, helping it understand context and meaning.