Natural Language Processing · Level 2 of 5
Perplexity
The exponential of average negative log probability under a sequence model.
Lower perplexity means higher assigned probability on comparable tokenized evaluation data.
Example
Two models using the same tokenizer are evaluated on held-out text.
Listen to the definition and example
Audio transcript
Perplexity. The exponential of average negative log probability under a sequence model. Lower perplexity means higher assigned probability on comparable tokenized evaluation data. For example: Two models using the same tokenizer are evaluated on held-out text.
Explore this concept
Why it matters
This helps you understand how language becomes data and where processing can lose meaning.
Start with
Related concepts
Quick recall question
Try answering before looking back at the definition.