Quiz: Benchmark Contamination and Why Public Scores Mislead

Benchmark contamination happens when test data or evaluation methods become compromised through repeated exposure or inappropriate use.

The most common mechanism of benchmark contamination involves data leakage where information from the target variable inadvertently enters the training data.

Public scores often misrepresent actual system performance because they measure against idealised conditions that rarely occur in practice.

A practical example of misleading public scores involves language models that perform well on factual queries but poorly on ambiguous requests or nuanced social situations.

Effective evaluation approaches should move beyond simple accuracy metrics towards more nuanced measures of system behaviour.