YangyiYY/rl-value-confidence-train-math

rl-value-confidence-train-math Math-domain training set (~10K) for a confidence / correctness estimator, sampled from the numina portion of PRIME-RL/Eurus-2-RL-Data. Pairs with YangyiYY/rl-value-eval-math (eval) and a 50K RL-training split from the same pool (disjoint). Sampled ~uniformly across the 6 numina sub-sources (cn_k12, synthetic_math, olympiads, synthetic_amc, aops_forum, amc_aime), deduplicated by problem text, and disjoint from the RL-train and eval splits (0… See the full description on the dataset page: https://huggingface.co/datasets/YangyiYY/rl-value-confidence-train-math.