Abbey4799/MetaR-Metaphorical-Riddle

METAR是一个隐喻谜题数据集，旨在支持元推理能力的研究。该数据集包含约3,444个样本，用于带有可验证奖励的强化学习（RLVR）。每个样本包含任务提示（instruction）、合成的隐喻谜题（riddle）、唯一正确答案（answer）以及三个层次分类（h1, h2, h3）。研究发现，隐喻推理是一种跨领域的基础技能，其效果与模型规模相关，且训练会增强模型的反思性思维。

METAR is a metaphorical riddle dataset designed to support research on meta-reasoning capabilities. The dataset contains approximately 3,444 samples for reinforcement learning with verifiable rewards (RLVR). Each sample includes a task prompt (instruction), a synthesized metaphorical riddle (riddle), a unique ground truth answer (answer), and three levels of categorization (h1, h2, h3). Research findings indicate that metaphor reasoning is a foundational skill that transfers across domains, its effectiveness is scale-dependent, and training enhances reflective thinking in models.