cot-oracle-eval-hinted-mcq
收藏资源简介:
该数据集名为 'CoT Oracle Eval: hinted_mcq',主要用于评估链式思维(CoT)模型在处理带有提示的多项选择题(MCQ)时的性能。数据来源于 GSM8K 测试集,包含 100 个训练样本。每个样本包含多个字段,如评估名称、唯一示例 ID、干净提示、测试提示、正确答案、提示答案以及多个元数据字段(如提示的准确性、提示的微妙程度、提示文本等)。此外,数据集还包含了预计算的 Qwen3-8B 模型在干净提示和测试提示上的链式思维响应和答案。数据集适用于评估模型在面对提示时的鲁棒性和准确性,特别是在数学推理任务中的应用。
This dataset, designated as 'CoT Oracle Eval: hinted_mcq', is primarily intended to assess the performance of Chain-of-Thought (CoT) models when solving hinted multiple-choice questions (MCQs). Derived from the GSM8K test set, it comprises 100 training samples. Each sample contains multiple fields, including the evaluation name, unique example ID, clean prompt, test prompt, correct answer, hinted answer, and several metadata fields (e.g., prompt accuracy, prompt subtlety, prompt text, etc.). Furthermore, the dataset includes pre-computed Chain-of-Thought responses and answers generated by the Qwen3-8B model on both clean prompts and test prompts. This dataset is applicable for evaluating the robustness and accuracy of models when encountering hints, especially for mathematical reasoning tasks.



