DataFlex-RL:RLVR 数据策略评估平台
原标题:Paper page - DataFlex-RL: An Evaluation Platform for RLVR Data Policies
AI 摘要
该论文提出 DataFlex-RL,一个用于在统一 GRPO 配方下比较 RLVR 数据策略(rollout 选择、重加权、领域混合)的评估平台。实验使用 Qwen2.5-7B-Base 和 Llama-3.1-8B-Base,在 12 个数学、逻辑与科学基准上测试 13 种配置、12 个匹配种子,发现均匀采样比未训练检查点平均准确率提升 7.76 个百分点,而没有任何替代策略相对均匀采样取得统计上可靠的改进。研究还发现基准组成会显著改变结论:用偏数学的六基准摘要重新评分时,排名与领域均衡的 12 基准摘要呈负相关(-0.33)。
正文节选
DataFlex-RL: An Evaluation Platform for RLVR Data Policies Abstract DataFlex-RL evaluates reinforcement learning data policies and finds that uniform sampling matches or exceeds adaptive rollout selection, reweighting, and domain mixing across math, logic, and science benchmarks. Data policies for reinforcement learning with verifiable rewards (RLVR) determine which rollouts are used, how strongly they are weighted, and which domains contribute to subsequent training batches. We introduce DataFl