CoT-Collection

# Dataset Card for Dataset Name

## Dataset Description

- **Homepage:https://github.com/kaistAI/CoT-Collection**
- **Repository:https://github.com/kaistAI/CoT-Collection**
- **Paper:https://arxiv.org/abs/2305.14045**
- **Point of Contact:[e-mail]**

### Dataset Summary


The CoT Collection is a dataset designed to induce Chain-of-Thought (CoT) capabilities into language models.
While proprietary LLMs excel at generating Chain-of-Thoughts based on prompting, smaller LMs do not have this capability. Thus, by fine-tuning to generate Chain-of-Thoughts, it could acquire such abilities.

The CoT Collection provides 1.84 million Chain-of-Thoughts augmented across 1060 tasks from the Flan Collection.\\
Experimental results show that fine-tuning on the CoT Collection results in (1) better zero-shot performance and (2) a better base model for few-shot learning.

We also provide a multilingual version of CoT Collection at this [link](https://huggingface.co/datasets/kaist-ai/Multilingual-CoT-Collection).

### Supported Tasks and Leaderboards

1060 tasks chosen from the Flan Collection.

The list of categories within the CoT Collection are:
* Natural Language Inference
* Extractive Question Answering
* Closed Book Question Answering
* Science
* Toxic Classification
* Arithmetic
* Program Execution
* Dialogue
* Ethics
* Commonsense Reasoning
* Multiple Choice Question Answering

### Languages

English

## Dataset Structure
* source: The input that is given to the language model (LM).
* target: The ground truth answer to the source.
* rationale: The Chain of Thought (CoT) that explains how the target could be derived from the source.
* task: A category that shows which dataset the source and target was extracted from.

In our paper, we trained the underlying language model to generate in the following format:
```
\{rationale\}
[RESULT]
\{target\}
```
Then during evaluation, we parsed the prediction after the phrase ```[RESULT]```.

### Data Splits

| name | train |
|-------------------|------:|
|CoT-Collection|1837928|

### Citation Information
If you find the following model helpful, please considering citing our paper!

```
@article{kim2023cot,
title={The CoT Collection: Improving Zero-shot and Few-shot Learning of Language Models via Chain-of-Thought Fine-Tuning},
author={Kim, Seungone and Joo, Se June and Kim, Doyoung and Jang, Joel and Ye, Seonghyeon and Shin, Jamin and Seo, Minjoon},
journal={arXiv preprint arXiv:2305.14045},
year={2023}
}
```

# 数据集卡片：数据集名称

## 数据集概述
- **主页：https://github.com/kaistAI/CoT-Collection**
- **代码仓库：https://github.com/kaistAI/CoT-Collection**
- **相关论文：https://arxiv.org/abs/2305.14045**
- **联系方式：[e-mail]**

### 数据集摘要

思维链（Chain-of-Thought，CoT）数据集（CoT Collection）是一款专为赋能语言模型习得思维链能力而打造的数据集。尽管闭源大语言模型（Large Language Model, LLM）凭借提示工程便可出色生成思维链，但小型语言模型并不具备此类能力。因此，通过针对思维链生成进行微调，小型模型即可习得此类技能。

CoT Collection 共包含来自Flan Collection的1060项任务，并为其扩增了184万条思维链样本。实验结果表明，在CoT Collection上进行微调可达成两大效果：(1) 更优异的零样本（Zero-shot）表现；(2) 适配少样本（Few-shot）学习的更优质基础模型。

我们还在此[链接](https://huggingface.co/datasets/kaist-ai/Multilingual-CoT-Collection)中提供了CoT Collection的多语言版本。

### 支持的任务与评测榜单
从Flan Collection中选取的1060项任务。

CoT Collection包含以下任务类别：
* 自然语言推理
* 抽取式问答
* 闭卷问答
* 科学推理
* 毒性分类
* 算术推理
* 程序执行
* 对话系统
* 伦理判断
* 常识推理
* 多项选择问答

### 语言
英语

## 数据集结构
* source：输入给语言模型（Language Model, LM）的原始提示文本
* target：对应source的标准答案
* rationale：用于阐释如何从source推导出target的思维链（Chain-of-Thought，CoT）
* task：用于标注source与target所属数据集类别的任务类别标签

在本研究的论文中，我们将基础语言模型的生成格式设定为如下形式：

{rationale}
[RESULT]
{target}

在评测阶段，我们会提取`[RESULT]`短语之后的内容作为模型预测结果。

### 数据划分
| 数据集名称 | 训练集样本数 |
|-------------------|------:|
|CoT-Collection|1837928|

### 引用说明
若本数据集对你的研究有所帮助，请考虑引用我们的论文！

@article{kim2023cot,
title={The CoT Collection: Improving Zero-shot and Few-shot Learning of Language Models via Chain-of-Thought Fine-Tuning},
author={Kim, Seungone and Joo, Se June and Kim, Doyoung and Jang, Joel and Ye, Seonghyeon and Shin, Jamin and Seo, Minjoon},
journal={arXiv preprint arXiv:2305.14045},
year={2023}
}