Your bookmarks

Loading your bookmarks

Latest updates

New courses, skills, book chapters, and site improvements will appear here.

Create free accountSign upDashboardDashboard

Benchmark contamination

  • Failures
  • Research term
4 steps

Benchmark contamination is when test questions, or close versions of them, end up in the data used to train, tune, or choose a model, so its score may reflect familiarity with the test.

Models learn from huge collections of web text, where public test questions get posted and copied, so they slip in unnoticed. Few developers publish their full training data, so outsiders can rarely measure the overlap. A model that saw a question with its answer during training can recall the answer, which the test scores like a reasoned one.

Where you’ll meet it

In a team meeting, a colleague proposes building the company's writing assistant on whichever model tops a popular leaderboard. That test's questions and answers have been online for three years, and a commenter under the leaderboard shows that, given half of a test question, the model types out the rest word for word.

What to do

Treat the score as weaker evidence than it looks, and suggest a short trial in its place. Collect twenty requests from your team's recent work, which have never been posted online, run the top two models on them, and compare the answers side by side.

Easy to mix up

A fresh set of test questions largely fixes contamination, since the model has nothing left to recall. Eval-maxxing can survive a fresh set, because habits that please the grader carry over.

Check yourself

The company behind a model adds a stack of math textbooks to its training data, and its score on a school-math test rises. Is the rise, on its own, a sign of contamination?

Go deeperResearchers look for shared text and for drops on newer questions

When the training data is available, researchers search it for long runs of words shared with test items. Otherwise they compare scores on questions published before and after the model's training cutoff, the date its training data ends, or on copies with new names and numbers. A sharp drop on those points to contamination.

Surveys report contamination risks across widely used benchmarks such as HumanEval (coding), MMLU (general knowledge), and GSM8K (school math). Some newer benchmarks keep their questions private, which slows leaks, though scoring a model hosted by the company behind it still sends each question to that company's servers.

Practice Modern AI

Or take the learning quiz on a mix of every set