xx18/Polaris-Composition-1323K

该数据集是论文《Composition-RL: Compose Your Verifiable Prompts for Reinforcement Learning of Large Language Models》中引入的数据集，用于解决强化学习过程中出现的太容易提示（通过率=1）问题。它通过自动组合多个可验证问题生成单个更难的提示，以保持训练信号的信息量。数据集包括多个子集，如评估集（AIME24、AIME25、BeyondAIME、IMO-AnswerBench、GPQA和MMLU-Pro）、主训练集（MATH-Composition-199K）、课程RL训练集（MATH-Composition-Depth3）、跨领域训练集（Physics-MATH-Composition-141K）以及基于Polaris53K构建的组合提示集（Polaris-Composition-1323K）。This repository contains datasets introduced in the paper [Composition-RL: Compose Your Verifiable Prompts for Reinforcement Learning of Large Language Models](https://huggingface.co/papers/2602.12036). Composition-RL is a data-efficient Reinforcement Learning with Verifiable Rewards (RLVR) approach that addresses the problem of too-easy prompts (pass-rate = 1) that occur during training. It automatically composes multiple verifiable problems into a single, harder verifiable prompt to maintain informative training signals throughout the RL process. The datasets include evaluation sets (AIME24, AIME25, BeyondAIME, IMO-AnswerBench, GPQA, and MMLU-Pro), main training set (MATH-Composition-199K), curriculum RL training set (MATH-Composition-Depth3), cross-domain training set (Physics-MATH-Composition-141K), and compositional prompts constructed from Polaris53K (Polaris-Composition-1323K).