DAPO-Math-17k contains 13 of 40 AMC 2023 test problems

BytedTsinghua-SIA/DAPO-Math-17k (the RL prompt set from the DAPO paper, 1.79M rows = ~17k prompts repeated) contains 13 of the 40 AMC23 problems used as the 'amc23' eval (math-ai/amc23) verbatim (normalised substring match on the full problem text), each repeated ~138 times; for 12 of them reward_model.ground_truth equals the AMC23 reference answer. No AIME 2024 problems were found, consistent with the DAPO paper's own AIME24 eval, but RL papers that train on DAPO-Math-17k and report AMC23 are partially training on the test set (12/40 = 30% of items with the correct reward). Remove those prompts or report AMC23 on the 27-28 clean items.