Benchmark Contamination

When data used to train a model overlaps, intentionally or not, with the questions or answers in an evaluation benchmark, inflating the model's measured performance without a genuine improvement in capability.