Perplexity
Domain: Evaluation
🎯 Motivation
How surprised the model is by the next token. Lower perplexity = better prediction = better model (on language tasks).
📋 Concrete Example
Perplexity on a text: 3.2 (good - model predicted most words correctly) vs 15.7 (poor - model was surprised by many words). Lower is better.
🔗 Analogy
A student’s confidence in predicting the next word in a sentence - ‘The cat sat on the…’ If they guess mat confidently, low perplexity. If they cannot guess, high perplexity.