Introduces DataFlex-RL, a platform for comparing rollout-selection and reweighting data policies used in reinforcement learning with verifiable rewards (RLVR) under a shared GRPO recipe. Across 13 configurations and 12 matched seeds on Qwen2.5-7B-Base, none of eight tested rollout-selection/reweighting methods or three adaptive mixtures statistically outperformed uniform sampling, and evaluation rankings were sensitive to benchmark-set composition (a math-heavy subset produced a -0.33 correlation with the full 12-benchmark ranking). The findings expose gaps in how RLVR data-policy claims are currently validated.
