CompactAI-O/Opus-4.6-Reasoning-2160x

Opus-4.6-Reasoning-2160x数据集包含2,160个由Claude Opus 4.6生成的高质量推理轨迹，涵盖数学、竞争编程、逻辑、科学和语言任务。每个例子包括完整的问题、扩展的思维链和最终解决方案，适用于监督微调、思维链蒸馏和推理能力转移到较小模型。数据集经过严格的质量过滤，去除了1,145个不符合要求的例子，保留了2,160个实质性、完整的推理演示。数据集的平均每行令牌数约为758，总完成令牌数为1,636,368，分为简单、中等和困难三个难度级别，以及数学/推理和代码/逻辑两个类别。

The Opus-4.6-Reasoning-2160x dataset contains 2,160 high-quality reasoning traces generated by Claude Opus 4.6, covering mathematics, competitive programming, logic, science, and language tasks. Each example includes the full problem, an extended chain-of-thought, and a final solution, making the dataset suitable for supervised fine-tuning, chain-of-thought distillation, and reasoning-capability transfer to smaller models. The dataset underwent rigorous quality filtering, removing 1,145 non-compliant examples and retaining 2,160 substantive, complete reasoning demonstrations. The dataset averages ~758 tokens per row, with a total of 1,636,368 completion tokens, and is categorized into easy, medium, and hard difficulty levels, as well as math/reasoning and code/logic categories.