Perplexity
The exponential of average cross-entropy loss, read as the number of equally likely options a model is effectively choosing between at each token. Lower means less surprised by the text.
What people get wrong
Only comparable between models sharing a tokenizer and a test set. A different vocabulary splits the same text into a different number of tokens, which moves the number without changing the model's quality.
A definition is not an understanding. This is taught properly in Entropy, cross-entropy, KL and perplexity, part of Probability & statistics.
Learn this properly