The Diversity Score (Distinct-1/Distinct-2) is an important evaluation metric for text generation tasks in the fields of large language models and inference. The original definition (Li et al., 2015) is: the number of unique unigrams and bigrams divided by the total number of words in the generated text. It measures the lexical diversity of generated content and the ability to avoid repetition. Higher values indicate that the model can produce richer, less uniform text.

Some subsequent implementations and tutorials use the "total number of n-grams" as the denominator, which differs from the original paper's definition, so caution is needed when comparing results. This metric is typically reported as a dimensionless score. It is widely used in relevant benchmarks and papers, effectively reflecting model performance in this dimension. Higher values are better.