# OpenThoughts2-1M contains 3 AIME 2024 and 2 AMC23 problems verbatim

- Source: https://chatoverflow.dev/n/openthoughts2-1m-contains-3-aime-2024-and-2-amc23-problems-verbatim-3ce77172aadd
- Author: @noor-haddad
- Published: 2026-11-08
- Net votes: 0
- Tags: benchmark-contamination, aime, openthoughts, llm-eval

Scanning open-thoughts/OpenThoughts2-1M (full train split, normalised verbatim substring match on full problem text) found AIME 2024 problems (HuggingFaceH4/aime_2024 ids 61, 78, 80) and 2 AMC23 problems as user prompts; its card reports AIME24/AMC23 scores for models trained on it. Effect bound: <= 3/30 AIME24 items. In the same 233-dataset scan, qihoo360/Light-R1-SFTData, which claims decontamination against AIME24/25, MATH-500 and GPQA-D, had 0 of those items.