PuzzleClone
收藏资源简介:
PuzzleClone是一个由HiThink Research和香港科技大学共同创建的高质量数学和逻辑数据集,包含超过83,657个经过验证的谜题,用于加强大型语言模型(LLMs)的推理能力。该数据集采用Satisfiability Modulo Theories (SMT)技术生成,每个谜题都经过编码、生成和验证三个阶段。数据集涵盖了广泛的难度和格式,为当前最先进的模型提出了重大挑战。通过PuzzleClone训练,可以提高模型在逻辑和数学基准测试中的表现,从而解决大型语言模型在逻辑推理方面的挑战。
PuzzleClone is a high-quality mathematical and logical dataset co-created by HiThink Research and The Hong Kong University of Science and Technology, containing over 83,657 verified puzzles designed to enhance the reasoning capabilities of Large Language Models (LLMs). This dataset is generated using Satisfiability Modulo Theories (SMT) technology, with each puzzle undergoing three stages: encoding, generation, and validation. It covers a wide range of difficulty levels and formats, posing significant challenges to current state-of-the-art models. Training with PuzzleClone can improve model performance on logical and mathematical benchmarks, thereby addressing the challenges of logical reasoning faced by Large Language Models.
- 1通过HiThink Research, 香港科技大学 · 2025年



