skipharlow/claude-opus-4.6-10000x

这是一个使用Claude Opus 4.6合成的高保真推理数据集，旨在捕获模型的内部“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。该数据集结合了高难度数学问题（如GSM8K和MATH）与通用逻辑谜题及多步骤指令，每行包含隐藏的推理轨迹，模型在提供最终答案前会进行内部“思考”。它主要用于监督微调（SFT）和蒸馏，以让较小的开源模型（如Qwen3.5系列）继承Claude Opus 4.6的复杂推理模式，从而学习过程导向的思维而非仅答案的模式匹配。训练此类数据有助于减少模型在非数学任务中的“幻觉”，提升逐步验证能力和跨领域泛化（如编码、法律分析和结构化写作），最终提高在BigBench Hard和GSM8K等基准测试上的性能。数据集格式为JSONL（带推理轨迹的对话式），主要类别包括数学、符号逻辑和通用问题解决。This is a high-fidelity reasoning dataset synthesized using Claude Opus 4.6, designed to capture the internal "Chain of Thought (CoT)" and reasoning trajectories of models, with a particular focus on mathematical accuracy and structured logical deduction. The dataset combines highly challenging mathematical problems (such as GSM8K and MATH) with general logical puzzles and multi-step instructions. Each line contains hidden reasoning trajectories, where the model performs internal "thinking" prior to delivering the final answer. It is primarily utilized for Supervised Fine-Tuning (SFT) and knowledge distillation, allowing smaller open-source models (e.g., the Qwen3.5 series) to inherit the complex reasoning patterns of Claude Opus 4.6, thus learning process-oriented thinking instead of answer-only pattern matching. Training with this dataset helps mitigate model "hallucinations" in non-mathematical tasks, improve step-by-step verification capabilities and cross-domain generalization (such as coding, legal analysis, and structured writing), and ultimately boost performance on benchmarks including BigBench Hard and GSM8K. The dataset follows the JSON Lines (JSONL) format as dialogue-style data with embedded reasoning traces, and its main categories cover mathematics, symbolic logic, and general problem-solving.