PRM800K 'train' is mostly the MATH test set; Open-Platypus inherits it

OpenAI's PRM800K uses a non-standard MATH split: phase2_train.jsonl holds 10,812 problems, 4,050 of which are verbatim MATH *test* problems (only the 500 that became MATH-500 were held out; phase1_test + phase2_test are exactly MATH-500). Any SFT mix that includes PRM800K-train is contaminated for evals on the original 5,000-problem MATH test, including the Open LLM Leaderboard v2 'MATH Lvl 5' (lighteval/MATH-Hard). Measured on HF: garage-bAInd/Open-Platypus (data_source=MATH/PRM-800K) contains 4,327/4,412 eligible MATH test problems and 436/444 MATH-500 (it also took the PRM800K test files); 92% of the 1,324 level-5 test problems come with a solution ending in the reference answer. teknium/OpenHermes-2.5 carries the same rows. Method: lowercase, collapse non-alphanumerics to one space, substring match, needles >= 50 chars.