himanshunakrani9/qwen-reasoning-samples-20260421_220553

3个通过Qwen/Qwen3.6-35B-A3B模型生成的合成推理示例，使用结构化提示设计以引发前沿水平（Opus 4.7类）的多阶段推理。每个示例都通过质量过滤器，要求推理过程遵循明确的6阶段结构（理解→分解→探索→执行→验证→反思），并包括独立的验证步骤。数据集的结构包括主题标签、难度等级、自包含的问题陈述、带有6个标记阶段的长形式思维链、使用不同方法的独立答案验证、最终明确答案以及2-4个常见错误方法及其解释。数据集旨在用于训练小型模型模仿结构化、经过验证的推理过程。

3 synthetic reasoning examples generated with Qwen/Qwen3.6-35B-A3B via vLLM, using a structured prompt designed to elicit frontier-level (Opus 4.7 class) multi-phase reasoning. Each example is gated through a quality filter that requires the reasoning trace to follow an explicit 6-phase structure (Understand → Decompose → Explore → Execute → Verify → Reflect) and to include an independent verification step. The dataset schema includes topic category label, difficulty tier, self-contained problem statement, long-form chain-of-thought with 6 labeled phases, independent check of the answer using a different method, final crisp answer, and 2-4 common wrong approaches with explanations. Intended for SFT / distillation data for training smaller models to imitate structured, verified reasoning traces.