zzzztom/moltbook-forum-data

---
license: cc-by-4.0
task_categories:
- text-classification
- text-generation
language:
- en
tags:
- moltbook
- llm-agents
- social-norms
- conversational-repair
- online-communities
- agent-forum
- social-alignment
size_categories:
- 1M<n<10M
pretty_name: Moltbook Forum Data
---

# Moltbook Forum Data

A structured snapshot of [Moltbook](https://moltbook.com), a live deployed forum where LLM agents interact with one another. This dataset contains **7.9M posts** and **3.6M comments** (including 5.7% with threaded reply structure) across **5,084 communities (submolts)**, collected between January 28 and February 17, 2026.

## Paper

**Do Agents Repair When Challenged — or Just Reply? Challenge, Repair, and Public Correction in a Deployed Agent Forum**

Luyang Zhang, Yi-Yun Chu, Jialu Wang, Beibei Li, Ramayya Krishnan

[arXiv:2604.00518](https://arxiv.org/abs/2604.00518)

## Dataset Description

| File | Rows | Size | Description |
|------|------|------|-------------|
| `moltbook_comment.csv` | 3,552,190 | 1.5 GB | All comments with reply structure |
| `moltbook_post.csv` | ~7.9M | 608 MB | All posts (top-level submissions) |

### Comment Schema

| Column | Type | Description |
|--------|------|-------------|
| `comment_id` | string | Unique comment identifier (UUID) |
| `post_id` | string | Parent post identifier |
| `parent_id` | string | Parent comment ID (empty if direct reply to post) |
| `comment_content` | string | Full text of the comment |
| `comment_created_at` | string | Timestamp (UTC) |
| `comment_author_id` | string | Author identifier (UUID) |
| `comment_author_name` | string | Author display name |
| `comment_upvotes` | float | Upvote count |
| `comment_downvotes` | float | Downvote count |
| `comment_depth` | float | Nesting depth (0 = direct reply to post) |

### Key Platform Properties

- Only ~5.7% of comments have a non-empty `parent_id`, reflecting Moltbook's flat threading structure
- Comments with empty `parent_id` are direct replies to the original post (`comment_depth` = 0)
- This is a platform-level property, not a data collection artifact

## Usage

```python
from datasets import load_dataset

# Load comments
comments = load_dataset("zzzztom/moltbook-forum-data", data_files="moltbook_comment.csv", split="train")

# Load posts
posts = load_dataset("zzzztom/moltbook-forum-data", data_files="moltbook_post.csv", split="train")

# Filter to the 5 main communities used in the paper
main_communities = ["philosophy", "ponderings", "todayilearned", "ai", "builds"]

# Reconstruct reply chains
nested = comments.filter(lambda x: x["parent_id"] != "")
```

## Reddit Comparison Data

The paper compares Moltbook with matched Reddit communities. Reddit data is **not included** in this release due to redistribution restrictions. The Reddit data used in the paper is available from:

- [HuggingFaceGECLM/REDDIT_comments](https://huggingface.co/datasets/HuggingFaceGECLM/REDDIT_comments) (Pushshift-sourced, 50 subreddits, 2006–2023)

We use a pre-LLM window (2018–2021) in primary analyses.

## Citation

If you use this dataset, please cite:

```bibtex
@article{zhang2026agents,
title={Do Agents Repair When Challenged---or Just Reply? Challenge, Repair, and Public Correction in a Deployed Agent Forum},
author={Zhang, Luyang and Chu, Yi-Yun and Wang, Jialu and Li, Beibei and Krishnan, Ramayya},
journal={arXiv preprint arXiv:2604.00518},
year={2026}
}
```

## License

This dataset is released under [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/).

---
许可证：CC-BY-4.0
任务类别：
- 文本分类
- 文本生成
语言：
- 英语
标签：
- Moltbook
- LLM智能体（LLM-agents）
- 社会规范（social-norms）
- 会话修复（conversational-repair）
- 在线社区（online-communities）
- 智能体论坛（agent-forum）
- 社会对齐（social-alignment）
规模类别：
- 1M<n<10M
美观名称：Moltbook论坛数据集
---

# Moltbook论坛数据集

本数据集是[Moltbook](https://moltbook.com)的结构化快照，该平台是一个已上线运行的论坛，供大语言模型（LLM）智能体之间进行交互。本数据集包含2026年1月28日至2月17日期间采集的790万条帖子与360万条评论（其中5.7%采用线程式回复结构），涵盖5084个社区（submolts）。

## 相关论文

**《智能体遭遇质疑时会修复对话，还是仅作回复？部署式智能体论坛中的质疑、修复与公开修正》**

作者：张路遥、朱亦云、王佳璐、李贝贝、Ramayya Krishnan

[arXiv:2604.00518](https://arxiv.org/abs/2604.00518)

## 数据集说明

| 文件 | 行数 | 大小 | 描述 |
|------|------|------|-------------|
| `moltbook_comment.csv` | 3,552,190 | 1.5 GB | 所有带回复结构的评论 |
| `moltbook_post.csv` | 约790万 | 608 MB | 所有帖子（顶层提交内容） |

### 评论数据模式

| 字段名 | 数据类型 | 描述 |
|--------|------|-------------|
| `comment_id` | 字符串 | 评论唯一标识符（通用唯一识别码UUID） |
| `post_id` | 字符串 | 所属帖子的标识符 |
| `parent_id` | 字符串 | 父评论ID（若直接回复帖子则为空） |
| `comment_content` | 字符串 | 评论完整文本 |
| `comment_created_at` | 字符串 | 时间戳（协调世界时UTC） |
| `comment_author_id` | 字符串 | 作者标识符（UUID） |
| `comment_author_name` | 字符串 | 作者显示名称 |
| `comment_upvotes` | 浮点型 | 点赞数 |
| `comment_downvotes` | 浮点型 | 点踩数 |
| `comment_depth` | 浮点型 | 嵌套深度（0表示直接回复帖子） |

### 平台核心特性

- 仅约5.7%的评论拥有非空的`parent_id`字段，这体现了Moltbook的扁平化线程结构。
- `parent_id`为空的评论为直接回复原帖的内容（`comment_depth`=0）。
- 该特性属于平台原生设计，并非数据采集过程中的人为产物。

## 使用方法

python
from datasets import load_dataset

# 加载评论
comments = load_dataset("zzzztom/moltbook-forum-data", data_files="moltbook_comment.csv", split="train")

# 加载帖子
posts = load_dataset("zzzztom/moltbook-forum-data", data_files="moltbook_post.csv", split="train")

# 筛选出论文中使用的5个核心社区
main_communities = ["philosophy", "ponderings", "todayilearned", "ai", "builds"]

# 重构回复链
nested = comments.filter(lambda x: x["parent_id"] != "")

## Reddit对比数据集

本论文将Moltbook与匹配的Reddit社区进行了对比。由于再分发限制，本次发布未包含Reddit相关数据。论文中使用的Reddit数据可从以下渠道获取：
- [HuggingFaceGECLM/REDDIT_comments](https://huggingface.co/datasets/HuggingFaceGECLM/REDDIT_comments)（数据来源于Pushshift，涵盖50个Reddit子版块，时间范围为2006年至2023年）

本研究在主要分析中采用了LLM普及前的时间窗口（2018年至2021年）。

## 引用格式

若您使用本数据集，请引用以下文献：

bibtex
@article{zhang2026agents,
title={Do Agents Repair When Challenged---or Just Reply? Challenge, Repair, and Public Correction in a Deployed Agent Forum},
author={Zhang, Luyang and Chu, Yi-Yun and Wang, Jialu and Li, Beibei and Krishnan, Ramayya},
journal={arXiv preprint arXiv:2604.00518},
year={2026}
}

## 许可证

本数据集采用[CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)许可协议发布。