saracandu/smol_reas_traces

该数据集包含在小型语言模型（SLMs）鲁棒性评估实验中生成的扩展思维链（Chain-of-Thought, CoT）推理轨迹。它详细展示了模型在不同文本扰动下如何处理数学推理任务。数据集包括多个配置，每个配置包含输入文本、16个独立的CoT推理路径、从轨迹中提取的答案、真实答案以及扰动类型（在plus配置中）。数据集还提供了模型在GSM8K标准基准和不同扰动类型下的性能比较。

This dataset contains extended Chain-of-Thought (CoT) reasoning traces generated during robustness evaluation experiments on Small Language Models (SLMs). It provides a detailed look at how models handle mathematical reasoning under various textual perturbations. The dataset includes multiple configurations, each containing input text, 16 independent CoT reasoning paths, extracted answers from the traces, ground truth answers, and perturbation types (in plus configurations). It also provides performance comparisons of models on the GSM8K standard benchmark and across different perturbation types.