The Perplexity Metric is... #flashcard - Intuition: "How surprised is the model by the actual text?" - Lower values indicate the model assigns higher probability to the correct tokens - Calculation: 2^(negative log likelihood per token) <!--ID: 1751507777225-->