What is perplexity and how interpret it?

Perplexity measures how surprised a language model is.
Jurafsky & Martin

How It Works:

Perplexity quantifies a model?s uncertainty over a text sequence: lower values mean the model predicts the next token more confidently.

Key Benefits:

  • Model comparison: Benchmark language models on the same data.
  • Hyperparameter tuning: Guide model size and regularization choices.
  • Progress tracking: Monitor learning during training.

Real-World Use Cases:

  • Language modeling: Evaluate GPT variants.
  • Speech recognition: Compare acoustic models.

FAQs

Is lower always better?
What's a good range?