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

Attempt it

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

Free · 2 AI evals / day

Quick reference

TL;DR

N-gram overlap scan detects verbatim memorization; canary strings detect training-data inclusion. Both have limits: n-grams miss paraphrases, canaries require pre-publication insertion.

Memory aid

Sign in for the mnemonic that makes this stick. See Pro pricing.

Easy to grasp

Imagine you wrote a secret exam and worry a student got a copy. Test 1: read the first half of each question aloud and see if the student can finish it word for word. If they can, they saw the exam. Test 2: before printing, hide a made-up word on the exam. After the test, ask students to write down every made-up word they remember. If anyone writes it, they saw the exam before test day.

Key concepts

Contamination Detection Eval Set Integrity

Detailed answer & concept explanation

~3 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.

Sign in for 5 free unlocks/day or go Pro for unlimited. See Pro pricing.

Quick scan · 2–4 min

Everything important, quickly.

Open with why contamination matters: it inflates scores and measures memorization. Walk the two tests: n-gram overlap scan with the completion test variant, then canary-based detection. Name the limitations of each. Close with the unsolved gap (paraphrase contamination) and the compensating strategy (dynamic eval sets).

Real-world examples

Real products, models, and research that use this idea.

OpenAI's GPT-4 technical report discusses n-gram overlap analysis for contamination detection across standard benchmarks.
The BigBench and HELM benchmark suites include contamination analysis sections that report n-gram overlap between eval items and known web corpora.
Anthropic uses canary-based probing alongside n-gram analysis in their internal evaluation pipelines to verify that held-out eval sets remain uncontaminated.
LiveBench rotates eval questions monthly specifically to stay ahead of contamination from model training cycles.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QA model scores 95% on your benchmark but only 70% on a fresh eval set covering the same skills. Is this proof of contamination?

Sign in or go Pro to see the approach for this follow-up.

QYou want to build a contamination-resistant benchmark. What design choices help?

Sign in or go Pro to see the approach for this follow-up.

Don't say this

Red flags and common mistakes that signal junior thinking. Click to expand.

Most common mistake

Relying solely on n-gram overlap, which misses paraphrased contamination where the model benefits from seeing a rephrased version of the eval question.

Sign in or go Pro to see all red flags and common mistakes. See Pro pricing.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

Sign in for 5 unlocks/day or go Pro for unlimited revision sheets. See Pro pricing.

References

Primary sources. Skim if you want the original framing.