{"id":42004,"date":"2026-01-30T13:49:46","date_gmt":"2026-01-30T12:49:46","guid":{"rendered":"https:\/\/csw.agency\/glossar\/perplexity-3\/"},"modified":"2026-01-30T13:49:46","modified_gmt":"2026-01-30T12:49:46","slug":"perplexity-3","status":"publish","type":"glossar","link":"https:\/\/csw.agency\/en\/glossar\/perplexity-3\/","title":{"rendered":"Perplexity"},"content":{"rendered":"

The term \u201ePerplexity<\/a>\u201c originates from information theory and the natural language processing (NLP)<\/a> and serves as a key metric for evaluating the performance of language models, including large language models (LLMs)<\/a>. It quantifies the uncertainty or \u201eastonishment\u201c of a model when predicting a sequence of words or sentences.<\/p>\n

What perplexity measures<\/h2>\n

Essentially, perplexity measures how well a probability model predicts a sample. For language models, a low perplexity value indicates that the model can predict the next sequence of words in a sequence with high confidence and accuracy. A high value, on the other hand, indicates that the model is less confident and has difficulty making accurate predictions. Perplexity can also be thought of as the number of equally likely options that the model considers on average for each prediction.<\/p>\n

Mathematically, perplexity is defined as the exponentiation of the cross entropy. The cross entropy measures how well a predicted distribution approximates the true distribution. A perplexity of 1 means that the model has no uncertainty and predicts the sequence perfectly. Values above 1 indicate a certain degree of uncertainty.<\/p>\n

Significance for AI development<\/h2>\n

Perplexity<\/a> is an essential tool for assessing the predictive power and overall performance of language models. It provides important insights for the development, evaluation and optimization of models. Developers use this metric to:<\/p>\n