euclaise/MegaCoT

MegaCoT数据集是对Flans CoT submix的扩展和重新创建，排除了ecqa数据集，并添加了多个额外的数据集，如ROPES、Entailmentbank、MATH等。该数据集包含rationale、target、source和prompt四个字段，主要用于训练和推理任务。数据集中的每个子集都遵循其原始数据集的许可协议。

The MegaCoT dataset is a slightly more detailed re-creation of the Flans CoT submix for reinforcement learning (RL), excluding the ecqa dataset which contains synthetic generations. It includes extra datasets not in the original CoT collection, such as ROPES, Entailmentbank (excluding task 2), MATH (up to level 2), cos_e, feasibilityQA, TAL-SCQ5K, Quartz, boolq, and FLUTE. The dataset features include rationale, target, source, and prompt, all of which are string types. The dataset is split into a training set containing 629,377 samples, occupying 189,154,865 bytes. The download size of the dataset is 90,889,666 bytes, and the dataset size is 189,154,865 bytes.

原始信息汇总

数据集概述

数据集信息

• 特征:
  • rationale: 类型为字符串
  • target: 类型为字符串
  • source: 类型为字符串
  • prompt: 类型为字符串
• 分割:
  • train: 字节数为189154865，样本数为629377
• 下载大小: 90889666字节
• 数据集大小: 189154865字节

配置

• 配置名称: default
  • 数据文件:
    • train: 路径为data/train-*

额外数据集

• ROPES
• Entailmentbank (排除任务2)
• MATH (最高到级别2)
• cos_e
• feasibilityQA
• TAL-SCQ5K
• Quartz
• boolq
• FLUTE

版权声明

• 数据集无特定版权，各子集遵循其原始数据集的许可协议。