Perplexity is a statistical measure used to evaluate how well a probabilistic model predicts a sample, commonly applied to language models. For a sequence of length N, it is defined as the Nth root of the inverse probability assigned to the sequence by the model, equivalently exp(−(1/N) sum log p(wi|context)). Lower perplexity indicates better predictive performance and is closely related to cross-entropy loss. It depends on the dataset and tokenization, and serves as a convenient, model-agnostic metric for comparing different models or configurations.

Playbooks

Perplexity

Measure of a model's uncertainty when predicting the next token.

‌

‌

‌

‌

Recent Views

Recent Views

‌

‌

‌