zabr946/url-quuen

这是一个使用Claude Opus 4.6合成的高保真推理数据集，旨在捕捉模型的内部“思维链”和推理轨迹，特别关注数学准确性和结构化逻辑演绎。数据集结合了高难度数学问题（如GSM8K、MATH）与通用逻辑谜题和多步骤指令，每行包含一个隐藏的推理轨迹，模型在提供最终答案前会“思考”问题。通过暴露微调模型于这些内部独白，模型学习过程导向思维，而不是仅依赖答案的模式匹配。该数据集适用于监督微调（SFT）和蒸馏，使较小的开源模型能够继承Claude Opus 4.6的复杂推理模式，从而提高在BigBench Hard和GSM8K等基准测试上的性能，而不增加参数数量。This is a high-fidelity reasoning dataset synthesized using Claude Opus 4.6, designed to capture the internal "chain of thought" and reasoning trajectories of models, with a particular focus on mathematical accuracy and structured logical deduction. The dataset combines high-difficulty mathematical problems (such as GSM8K and MATH) with general logical puzzles and multi-step instructions. Each line contains a hidden reasoning trajectory, where the model "thinks" through the problem before providing the final answer. By exposing fine-tuned models to these internal monologues, the model's learning process is oriented toward reasoning rather than relying solely on pattern matching of final answers. This dataset is suitable for supervised fine-tuning (SFT) and knowledge distillation, enabling smaller open-source models to inherit the complex reasoning patterns of Claude Opus 4.6, thereby improving their performance on benchmark tests including BigBench Hard and GSM8K without increasing the number of model parameters.