BIG-bench task designed to detect evidence that a language model was trained on benchmark data.
unassessed
| Category | reasoning |
|---|---|
| Subcategory | data contamination detection |
| Page status | active |
| Metric | log10 p-value of canary probability deviation (log10_p_dev) |
| Direction | higher_is_better |
| Dataset licence | Apache-2.0 |
| Publisher | Google BIG-bench |
The task measures whether a model assigns anomalous conditional log-probability to a hard-coded BIG-bench canary GUID and to BIG-bench's own git commit hashes, compared with random control strings, as evidence the model was trained on BIG-bench's public repository.
Conditional log-probability scoring of fixed canary strings against random control strings; the model is not asked to generate text.
No model card in ModelSpec reports this benchmark yet.