ДалееСледующее руководство
Тесты искусственного интеллекта
Технический
РУКОВОДСТВО ПО ОСНОВАМ
Эталонное загрязнение
Benchmark contamination happens when questions or answers from an evaluation benchmark end up in a model's training data, so a high score can reflect memorization rather than real ability.
На этой странице4 минуты чтения
Обзор
It matters because leaderboard numbers drive research claims and buying decisions, and contaminated scores overstate how a model will do on problems it has never seen.
Глубокое погружение
Large language models are trained on web-scale text that includes GitHub, arXiv papers, forums and tutorial sites. Popular benchmarks are public by design, so their questions, answers and discussions of them get copied across the web and swept into training data. Contamination can be direct, where the exact test item appears, or indirect, through paraphrases, translations, solution write-ups, or synthetic data generated by another model that had seen the benchmark. It can also enter during fine-tuning if instruction datasets were assembled from benchmark-like sources. A useful distinction is between seeing only the question and seeing the question with its answer. The second is more damaging, because the model can recall the answer instead of working it out. The result is an inflated score that does not transfer to new problems. Lab awareness is not new. The GPT-3 paper (2020) ran a 13-gram overlap analysis between its benchmarks and training data and acknowledged that a bug left some overlaps in place. When training data is available, overlap search is the most direct check. When it is not, researchers use indirect methods: prompting the model with the start of a test item to see if it completes the rest verbatim; membership inference methods such as Min-K% Prob (Shi and colleagues, 2023), which look at how confidently a model predicts a text's tokens; comparing scores on original items against rewritten or newly written equivalents; and time-based splits that compare problems created before and after the training cutoff. Two misconceptions are common. First, overlap does not always raise scores much; some studies find modest effects on certain tasks. Second, finding no exact overlap does not prove a benchmark is clean, because paraphrased or translated copies slip past string matching. Mitigations include private held-out test sets, canary strings, regularly refreshed benchmarks such as LiveBench and LiveCodeBench, decontamination filters, and publishing contamination analyses alongside results.
Стратегическое воздействие
Более четкие решения
Это поможет вам отделить четкие технические заявления от маркетингового языка.
Стоимость и бюджет
Вы можете задать более эффективные вопросы по реализации, прежде чем тратить деньги или время.
Команда и рабочий процесс
Команды с общим пониманием принимают более эффективные решения по продуктам, политике и обучению.
The Future of Benchmark Contamination
Evaluation practice is shifting toward private or partly private test sets, benchmarks that add new items over time, and third-party evaluators who control the test data. That creates a real tension with openness, since public benchmarks are easy to reproduce and inspect. Greater transparency about training data would make contamination checks far easier, but many model developers do not publish full data details. Readers should treat any single static benchmark score with caution and look for results on fresh or held-out data before drawing strong conclusions.
Реальная реализация
A math benchmark's problems are reposted on forums with worked solutions; a web crawl collects those pages, and a model trained on it later reproduces the exact answers, including a quirk in the original wording.
Researchers at Scale AI wrote GSM1k, new grade-school math problems matched to GSM8K in style and difficulty, and found some models scored noticeably lower on the fresh set, a sign of overfitting to the public benchmark.
A coding benchmark that dates each problem shows a model solving far more problems published before its training cutoff than after, which is the pattern LiveCodeBench was designed to expose.
BIG-bench tasks include a unique canary string so data teams can filter any document containing it out of training corpora and later check whether a model can reproduce it.
Риски и ограничения
Разные команды могут использовать один и тот же термин по-разному, поэтому заранее определите масштаб.
Тесты могут выглядеть сильными, в то время как реальная производительность неравномерна.
Игнорирование качества данных и планов оценки часто приводит к нестабильным результатам.
Дорожная карта реализации
Начните с простого определения желаемого результата.
Перед тестированием выберите один показатель успеха и одно условие отказа.
Запустите небольшой пилотный проект с репрезентативными данными, а не отточенный демонстрационный набор.
Document where Benchmark Contamination helps and where simpler methods are better.
Продолжайте исследовать
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Benchmark Contamination quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Часто задаваемые вопросы
What is Benchmark Contamination?
Benchmark contamination happens when questions or answers from an evaluation benchmark end up in a model's training data, so a high score can reflect memorization rather than real ability. It matters because leaderboard numbers drive research claims and buying decisions, and contaminated scores overstate how a model will do on problems it has never seen.
What is benchmark contamination?
Contamination means the model may have seen test material during training, so its score can reflect memorization.
What is the purpose of the canary string included in BIG-bench tasks?
A unique identifier makes it easy to find and remove benchmark text from corpora and to probe whether a model has memorized it.
What did the GSM1k study suggest?
A drop on new problems matched in style and difficulty points to overfitting to the public benchmark rather than general math skill.
What overlap analysis did the GPT-3 paper run?
GPT-3's authors checked 13-gram overlap and acknowledged a bug that left some overlaps unremoved.
What is the core idea of Min-K% Prob?
The method averages the lowest k percent of token log-probabilities; memorized text scores unusually high because little in it surprises the model.
Продолжайте учиться
Связанные руководства
Другие руководства, выбранные по этой теме