The standard score for how well a language model predicts text: roughly, how surprised it is by what actually comes next. Lower is better. It is useful for comparing training runs within one model family, but it says little about whether a model is helpful, truthful or pleasant to use, which is why nobody chooses a product on it.
Technical view
The exponential of the average negative log likelihood per token, interpretable as the effective number of equally likely choices the model is deciding between at each step. Values are only comparable across models sharing a tokeniser and evaluation corpus, since a different subword vocabulary changes the denominator. It correlates with downstream capability within a family but breaks down across families and after instruction tuning, which typically raises perplexity on raw text while improving usefulness.
Related
Part of AI Terminology Index, section: Measuring Performance.