You suspect your custom eval set leaked into a model's training data. Describe two practical tests you would run to detect contamination.

When an eval set leaks into training data, benchmark scores measure memorization rather than real capability. Two practical tests detect this. **Test 1: N-gram overlap scan.** Feed the model the first half of each eval question and check if it completes the rest with unusual accuracy. Also compute 8-gram or 13-gram overlap between eval items and model outputs. If the model can reproduce eval text from prefixes, those items were memorized. **Test 2: Canary strings.** Before publishing the eval set, insert unique synthetic tokens that do not exist anywhere else. After training, test if the model can reproduce them. If it can, the eval set was in the training data. Both have limits. N-gram overlap misses paraphrased versions. Canaries must be inserted before training data collection, so they only work for eval sets you control from the start.

**N-gram overlap scan.** Compute exact 8-gram or 13-gram overlap between eval set items and model completions. The completion test is the practical version: feed the first half of each eval question to the model and measure how accurately it completes the second half. Suspiciously high completion accuracy signals memorization. Use 13+ grams to reduce false positives from common phrases. Limitation: n-gram overlap detects verbatim memorization but misses paraphrased contamination. A training set that includes rephrased versions of eval questions produces zero n-gram overlap while still inflating scores. **Canary-based detection.** Insert unique synthetic strings (UUID-like tokens, fictional entities, fabricated statistics) into the eval set before publication. After training, test whether the model reproduces these strings. Reproduction proves the eval set was in training data. Limitation: canaries must be inserted before training data collection. For already-published benchmarks, retroactive canary insertion is impossible. **Why this matters.** Contaminated eval scores measure memorization, not generalization. A model that scores 95% because it memorized the eval set may perform at 70% on genuinely novel inputs. Without contamination detection, you cannot trust benchmark comparisons.