Perplexity

A measure of how well a language model predicts a sample of text, specifically how surprised the model is by the text. Lower perplexity means the model finds the text more predictable and is therefore a better fit for that language. It is commonly used to compare language models but does not always correlate with performance on downstream tasks.