Kwai-Klear/RLEP_dataset

该数据集包含了用于论文RLEP: 基于经验回放的LLM推理强化学习中的实验数据。数据集遵循DAPO schema，分为测试集和训练集，其中测试集包含了AIME-2024、AIME-2025和AMC-2023数据，而训练集则包含了17k个数学问题，每个问题都有经验池存储在reward_model.candidates字段中。数据集的设计旨在优化数学推理任务的模型学习过程，通过回放高质量样本来避免无效探索，并专注于有希望的推理路径。

This dataset contains the experimental data used in the paper RLEP: Reinforcement Learning with Experience Replay for LLM Reasoning. The dataset follows the DAPO schema and is divided into a test set and a training set, with the test set including AIME-2024, AIME-2025, and AMC-2023 data, and the training set containing 17k math questions, each with an experience pool stored in the reward_model.candidates field. The design of the dataset aims to optimize the model learning process for math reasoning tasks by replaying high-quality examples to avoid fruitless exploration and focus on promising reasoning paths.