Training Data Contamination
A significant issue in AI evaluation referring to the presence of benchmark questions or their solutions within a model's training corpus. This causes high benchmark scores to reflect memorization rather than genuine reasoning ability, thereby undermining the validity and reliability of evaluation results.
Core Facts
Benchmarks face contamination risks; if training data contains test questions or approximate variants, a model's high scores may partly stem from memorization rather than genuine capability.
Benchmark data contamination refers to the phenomenon where, after a test set is made public, its questions may appear in the pre-training corpora of subsequent models, causing scores to reflect memorization rather than genuine reasoning ability.
Data poisoning attacks involve mixing malicious samples into training datasets, causing the model to exhibit incorrect behavior toward specific inputs after deployment.
Timeline (last 90 days)
Training data contamination is a common problem in machine learning, referring to a model being exposed during training to data that should not have been included, which may cause the model to learn to imitate the response patterns of other models.
PsyAgentBench employs a multi-dimensional factorial crossed design—including named vs. blind testing, standard vs. counterfactual versions, and persona manipulation—to guard against training data contamination.
Benchmark Gaming manifests at the technical level in forms such as training data contamination, targeted hyperparameter tuning, and output format pandering.
Data poisoning attacks involve mixing malicious samples into training datasets, causing the model to exhibit incorrect behavior toward specific inputs after deployment.
All Facts (6)
Benchmarks face contamination risks; if training data contains test questions or approximate variants, a model's high scores may partly stem from memorization rather than genuine capability.
90%VerifiedBenchmark data contamination refers to the phenomenon where, after a test set is made public, its questions may appear in the pre-training corpora of subsequent models, causing scores to reflect memorization rather than genuine reasoning ability.
80%VerifiedData poisoning attacks involve mixing malicious samples into training datasets, causing the model to exhibit incorrect behavior toward specific inputs after deployment.
65%UnverifiedTraining data contamination is a common problem in machine learning, referring to a model being exposed during training to data that should not have been included, which may cause the model to learn to imitate the response patterns of other models.
50%UnverifiedPsyAgentBench employs a multi-dimensional factorial crossed design—including named vs. blind testing, standard vs. counterfactual versions, and persona manipulation—to guard against training data contamination.
50%UnverifiedBenchmark Gaming manifests at the technical level in forms such as training data contamination, targeted hyperparameter tuning, and output format pandering.
50%