zen-E/GSM8k-Aug

该数据集名为GSM8k-Aug，用于训练CODI模型。数据集由两部分组成：GSM8k-Aug和GSM8k-Aug-NL。GSM8k-Aug通过提示GPT-4扩展了原始的GSM8k训练集至385k样本，并移除了所有自然语言交错的隐式CoT，仅保留了结构化的数学表达式。GSM8k-Aug-NL则保留了自然语言解释，用于评估方法的泛化能力和压缩冗长CoT的有效性。数据集的格式为：问题||CoT####答案。

This dataset is named GSM8k-Aug, used for training the CODI model. The dataset consists of two parts: GSM8k-Aug and GSM8k-Aug-NL. GSM8k-Aug extends the original GSM8k training set to 385k samples by prompting GPT-4 and removes all interleaved natural language within the implicit CoT, leaving only structured mathematical expressions. GSM8k-Aug-NL preserves natural language explanations to assess the generalization ability and the effectiveness of the approach in compressing verbose CoTs. The format of the dataset is: <question>||<CoT>####<Answer>.