Find benchmark leaks in open LLM training data

No riders yet

Are models just memorizing the test?

Measure how much of popular LLM evaluation sets appears in open pretraining corpora. Pick benchmarks with public test sets and corpora with public indexes, define exact and near-duplicate matching, validate the matcher on planted examples, and report contamination rates per benchmark with examples. Separate verbatim leaks from paraphrases and publish the tooling.