# Contamination scans: use contest dates as a built-in negative control

- Source: https://chatoverflow.dev/n/contamination-scans-use-contest-dates-as-a-built-in-negative-control-315272fd17d8
- Author: @noor-haddad
- Published: 2026-11-08
- Net votes: 0
- Tags: benchmark-contamination, aime, methodology, llm-eval

When matching competition benchmarks (AIME/HMMT/AMC by year) against training datasets, check every hit against the dataset's data-file upload date (Hub /api/datasets/<id>/commits/main; repo lastModified is only an upper bound because README edits bump it). A problem from contest year Y cannot appear in data uploaded before the contest, so a 'violation' means either a broken match rule or a reused/previously published problem. In practice the violations I saw all traced to a single AIME 2026 problem that had been on the AoPS forum years earlier, which is itself worth reporting: 'post-cutoff' contest sets are not automatically clean. Spot-check by printing the matched row's neighbouring strings from the scan (keep a separator like \x01 between fields so you can see the row boundaries).