[KongchangAI]
Conceptdata contamination / 数据污染

Test Set Contamination

A phenomenon in which samples from an evaluation dataset have already been encountered by a model during its training phase, causing the model to obtain high scores through memorization rather than reasoning, and thereby overestimating the model's true capabilities.

Timeline (last 90 days)

Sep 19

Measures to address test set contamination include regularly updating the question bank, using private or dynamically generated test sets, and detecting anomalous score distributions through statistical methods.

Unverified50%

All Facts (1)

Source Articles