Evaluation questions, answers, or close variants can leak through public corpora, generated data, or iterative prompt tuning. Deduplication, provenance tracking, and audits reduce the risk but may not detect every semantic overlap.
Data contamination occurs when information that should be held out from training or development appears in data used to build a model.
Evaluation questions, answers, or close variants can leak through public corpora, generated data, or iterative prompt tuning. Deduplication, provenance tracking, and audits reduce the risk but may not detect every semantic overlap.