frasalvi/debategpt

---
license: cc-by-sa-4.0
language:
- en
tags:
- debates
size_categories:
- n<1K
---
# Dataset Card for DebateGPT

The DebateGPT dataset contains debates between humans and GPT-4, along with sociodemographic information about human participants and their agreement scores before and after the debates.
This dataset was created for research on measuring the persuasiveness of language models and the impact of personalization, as described in this pre-print: [On the Conversational Persuasiveness of Large Language Models: A Randomized Controlled Trial
](https://huggingface.co/papers/2403.14380).

## Dataset Details

The dataset consists of a CSV file with the following columns:

- **debateID**: ID of the debate. For *Human-Human* debates, the two participants taking part in the debate can be paired using this column.
- **treatmentType**: Treatment condition to which the debate belongs. One of *Human-Human*, *Human-AI*, *Human-Human, personalized*, *Human-AI, personalized*. In "personalized" conditions, participants' personal information is available to their opponents.
- **topic**: The proposition about which the debate is held.
- **gender**: The participant's gender.
- **age**: The participant's age group.
- **ethnicity**: The participant's age group.
- **education**: The participant's education level.
- **employmentStatus**: The participant's employment status.
- **politicalAffiliation**: The participant's political orientation.
- **side**: The side assigned to the participant in the debate (PRO or CON).
- **agreementPreTreatment**: The participant's agreement with the debate proposition, before the debate, on a 1-5 Likert scale (A^pre).
- **agreementPostTreatment**: The participant's agreement with the debate proposition, after the debate, on a 1-5 Likert scale (A^post).
- **sideAgreementPreTreatment**: The participant's agreement with the side opposing the one they were assigned to (i.e. their agreement with their opponent), before the debate (Ã^pre).
- **sideAgreementPostTreatment**: The participant's agreement with the side opposing the one they were assigned to (i.e. their agreement with their opponent), after the debate (Ã^post).
- **topicPrior**: The participant's prior exposure to the debate topic, on a 1-5 Likert scale.
- **argument**: The participant's argument.
- **rebuttal**: The participant's rebuttal.
- **conclusion**: The participant's conclusion.
- **argumentOpponent**: The opponent's argument.
- **rebuttalOpponent**: The opponent's rebuttal.
- **conclusionOpponent**: The opponent's conclusion.
- **perceivedOpponent**: The participant's perception of their opponent's identity (human or ai).

## Usage
```python
from datasets import load_dataset
dataset = load_dataset("frasalvi/debategpt")
```

## Citation
If you would like to cite our work or data, you may use the following bibtex citation:

```
@misc{salvi2024conversational,
title={On the Conversational Persuasiveness of Large Language Models: A Randomized Controlled Trial},
author={Francesco Salvi and Manoel Horta Ribeiro and Riccardo Gallotti and Robert West},
year={2024},
eprint={2403.14380},
archivePrefix={arXiv},
primaryClass={cs.CY},
doi = {10.48550/arXiv.2403.14380}
}
```

原始信息汇总

数据集概述：DebateGPT

数据集内容

• 描述: DebateGPT 数据集包含人类与GPT-4之间的辩论，以及关于人类参与者的社会人口统计信息和他们在辩论前后的同意分数。
• 目的: 用于研究语言模型的说服力和个性化影响。

数据集结构

• 文件类型: CSV
• 列信息:
  • debateID: 辩论ID
  • treatmentType: 辩论所属的治疗条件，包括 Human-Human, Human-AI, Human-Human, personalized, Human-AI, personalized
  • topic: 辩论主题
  • gender: 参与者性别
  • age: 参与者年龄组
  • ethnicity: 参与者种族
  • education: 参与者教育水平
  • employmentStatus: 参与者就业状态
  • politicalAffiliation: 参与者政治倾向
  • side: 参与者在辩论中的立场（PRO或CON）
  • agreementPreTreatment: 辩论前参与者对辩论命题的同意程度（1-5 Likert量表）
  • agreementPostTreatment: 辩论后参与者对辩论命题的同意程度（1-5 Likert量表）
  • sideAgreementPreTreatment: 辩论前参与者对其对手立场的同意程度
  • sideAgreementPostTreatment: 辩论后参与者对其对手立场的同意程度
  • topicPrior: 参与者对辩论主题的先前接触程度（1-5 Likert量表）
  • argument: 参与者的论点
  • rebuttal: 参与者的反驳
  • conclusion: 参与者的结论
  • argumentOpponent: 对手的论点
  • rebuttalOpponent: 对手的反驳
  • conclusionOpponent: 对手的结论
  • perceivedOpponent: 参与者对其对手身份的感知（人类或AI）

许可

• 类型: CC-BY-SA-4.0