---
license: mit
task_categories:
- question-answering
language:
- en
- zh
tags:
- human-behavior
- social-media
size_categories:
- 10K<n<100K
---


<p align="center">
  <img center src="https://i.postimg.cc/NjXSwQvY/FineRob.png" width = "150" alt="logo">
</p>
<h2 align="center">FineRob - Fine-Grained Social Media Behavior Simulation Dataset</h2>

## Paper
https://arxiv.org/abs/2412.03148

## Github
https://github.com/linkseed18612254945/FineRob

## Table of Contents

- [Overview](#overview)
- [Dataset Description](#dataset-description)
- [Download](#Download)
- [Introduction](#Introduction)
  
## Overview
**Finerob** is a novel fine-grained user behavior simulation dataset collected from three social media platform: X(Twitter), Reddit, Zhihu.
The dataset is design to evalute the role-play capacity of LLMs through three differnet action elements simulation.

<p align="center">
  <img center src="https://i.postimg.cc/52JHnQGw/RESA-Example-new.png" alt="example">
</p>

## Dataset Description

- **Name**: Finerob
- **Purpose**: Social media user behavior simulation by role-playing LLMs.
- **Form**: multiple-choice QA.
- **Size**: 78.6k fine-grained behavior QA records from 1,866 users.
- **Language**: English, Chinese.


## Introduction
We collect 78.6k fine-grained user behavior prediction QA data from 1866 users on social medias(including Twitter, Reddit and Zhihu)

You can find the detailed user infomation in *platform_user_info.json* through the *user_index*. 

*history_index* means the behavior index in user's complete timelines, basicly, you should predict this behavior by the history. The history_index  0 means it is the first behavior of the user.