Perplexity (Metric)
Definition
Perplexity measures a model's uncertainty about a sequence.
An exponential measure of the average uncertainty of a probabilistic model over a sequence; lower values are directly comparable only with the same tokenizer and dataset.
Simply explained
It shows how uncertain a model is about the considered text; comparisons require the same data basis and split.
Why it matters in practice
The term helps to select, develop, evaluate AI systems in a well-founded manner and realistically assess their technical risks.
Example from business
Two versions of the same language model are compared on the same test corpus.
Limits and pitfalls
- Comparing values across tokenizers or corpora or equating low perplexity with good factual accuracy.
Origin & history
Perplexity originates from information theory and language model evaluation.
Use in mid-sized companies
- Possible application: Two versions of the same language model are compared on the same test corpus.
Frequently asked questions
What does perplexity mean as a model metric?
An exponential measure of the average uncertainty of a probabilistic model over a sequence; lower values are directly comparable only with the same tokenizer and dataset. What mistakes should be avoided? Comparing values across tokenizers or corpora or equating low perplexity with good factual accuracy.
Sources
Source verified on 10.10.2026. Editorial application examples are hypothetical; no evidence of real customer results.
Editorially reviewed on: