reasoning-core/rc0

Reasoning Core是一个基于文本的RLVR（强化学习可验证奖励）环境，专为大型语言模型（LLMs）的符号推理训练设计。它涵盖了广泛的符号任务，包括完整的一阶逻辑（FOL）、使用TPTP的形式数学、新颖领域的正式规划以及语法任务。该数据集通过高通用性问题分布、外部工具验证和连续难度控制等设计原则，提供了几乎无限的训练实例。初步的零样本评估显示，Reasoning Core的任务对前沿LLMs具有挑战性，是提升未来模型推理能力的有力资源。

Reasoning Core is a text-based RLVR (Reinforcement Learning with Verifiable Rewards) environment designed for symbolic reasoning training in Large Language Models (LLMs). It encompasses a wide range of symbolic tasks, including full-fledged first-order logic (FOL), formal mathematics using TPTP, formal planning with novel domains, and syntax tasks. The dataset is built on key design principles such as high-generality problem distributions, verification via external tools, and continuous difficulty control, providing a virtually infinite supply of novel training instances. Initial zero-shot evaluations with frontier LLMs confirm the difficulty of Reasoning Cores tasks, positioning it as a promising resource to improve the reasoning capabilities of future models.