What is benchmark contamination for a language model?
Test questions leaked into the training data, so scores look better than real skill.
- A virus that corrupts the scores
- Users typing rude prompts
- Test questions slipping into the training data ✔
Why: Web-scale training sets can include copies of benchmark questions and answers. A model that saw them during training may have memorized them, which makes the benchmark score unreliable.
Huh, didn't know that: The GPT-3 paper admits a filtering bug left some test overlaps in, and the model was too costly to retrain.
Benchmark contamination is like acing a test because you saw the answer key last week. Impressive, but not learning.
References
Answer: Test questions slipping into the training data. Web-scale training sets can include copies of benchmark questions and answers. A model that saw them during training may have memorized them, which makes the benchmark score unreliable.
- Brown et al., Language Models are Few-Shot Learners (arXiv 2005.14165) arxiv.org/pdf/2005.14165
potential contamination of downstream tasks by having their test or development sets inadvertently seen during pre-training
Checked 2026-10-10. - Yang et al., Rethinking Benchmark and Contamination for Language Models with Rephrased Samples (arXiv 2311.04850) arxiv.org/abs/2311.04850
simple variations of test data (e.g., paraphrasing, translation) can easily bypass these decontamination measures
Checked 2026-10-10.
Think this answer is wrong? Tap "Challenge this answer" on the card in the game and tell us why.
More stuff everyone should know
That's stuff everyone should know.