N NexChron

Perplexity

Definition A statistical measure of how well a language model predicts a sample of text — lower perplexity means the model assigns higher probability to the actual next tokens and is therefore a better fit. Perplexity is a standard intrinsic evaluation metric for language models.

In Depth

Perplexity is computed as the exponent of the average negative log-likelihood per token. It is used to compare models trained on the same data distribution and to track training progress. While perplexity correlates loosely with downstream task performance, it does not capture instruction-following ability, factual accuracy, or safety — which is why benchmark suites and human evaluation remain essential complements.

Browse more terms

AI Agent AI Alignment Agentic AI Artificial General Intelligence Artificial Superintelligence Attention Mechanism AutoML Benchmark Bias (AI) Chain-of-Thought Computer Vision Deepfake Diffusion Model Edge AI Embedding Federated Learning Few-Shot Learning Fine-Tuning Function Calling GPT