README.md

---
license: cc-by-sa-4.0
task_categories:
- text-classification
language:
- ko
pretty_name: Youtube Bot Comment
size_categories:
- 100K<n<1M
dataset_info:
  features:
  - name: comment_id
    dtype: string
  - name: content
    dtype: string
  - name: author_name
    dtype: string
  - name: author_image_url
    dtype: string
  - name: video_id
    dtype: string
  - name: parent_id
    dtype: string
  - name: is_bot_comment
    dtype: bool
  splits:
  - name: train
    num_bytes: 39287858
    num_examples: 130080
  - name: test
    num_bytes: 11008299
    num_examples: 37167
  - name: validation
    num_bytes: 5589340
    num_examples: 18583
  download_size: 29190712
  dataset_size: 55885497
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
  - split: validation
    path: data/validation-*
---

> Note for users, README.md generated by Claude 4 Sonnet.

# Dataset Card for YouTube Korean Bot Comment Dataset

This dataset contains YouTube comments from the top 50 Korean videos, classified to identify bot-generated content for research in automated content detection and Korean natural language processing.

## Dataset Details

### Dataset Description

This dataset consists of 185,830 YouTube comments collected from the top 50 Korean videos and classified for bot detection research. Comments were extracted using the YouTube Data API v3 and labeled using Google's Gemini 2.5 Flash-Lite model to distinguish between human-generated and bot-generated content. The dataset serves as a resource for training and evaluating bot detection systems specifically for Korean social media content.

- **Curated by:** [MisileLab](https://huggingface.co/MisileLab)
- **Language(s) (NLP):** Korean (ko)
- **License:** CC BY-SA 4.0

### Dataset Sources

- **Repository:** YouTube Data API v3
- **Classification Model:** Google Gemini 2.5 Flash-Lite

## Uses

### Direct Use

This dataset is suitable for:
- Training bot detection models for Korean social media content
- Research on automated content generation patterns in Korean language
- Benchmarking natural language processing models for Korean text classification
- Social media authenticity analysis and content moderation research
- Comparative studies of bot behavior across different platforms

### Out-of-Scope Use

This dataset should not be used for:
- Generating or creating bot comments
- Harassment or targeting of specific YouTube users
- Commercial content moderation without proper validation
- Training models to evade bot detection systems
- Analysis outside the Korean language context without domain adaptation

## Dataset Structure

The dataset contains the following fields:

- `comment_id`: Unique identifier for each YouTube comment
- `content`: The actual text content of the comment in Korean
- `author_name`: Username of the comment author
- `author_image_url`: Profile image URL of the comment author
- `video_id`: Identifier of the YouTube video where the comment was posted
- `parent_id`: Identifier of parent comment (for replies, null for top-level comments)
- `is_bot_comment`: Boolean label indicating whether the comment is classified as bot-generated (True) or human-generated (False)

**Split Information:**
- Training set: 130,080 examples (70%)
- Test set: 37,167 examples (20%)
- Validation set: 18,583 examples (10%)

## Dataset Creation

### Curation Rationale

This dataset was created to address the lack of Korean-language resources for bot detection research. With the increasing prevalence of automated content on social media platforms, there is a critical need for language-specific datasets to train and evaluate detection systems for non-English content, particularly Korean.

### Source Data

#### Data Collection and Processing

1. **Video Selection**: The top 50 Korean videos were identified using YouTube's trending metrics and popularity indicators
2. **Comment Extraction**: Comments were systematically collected using the YouTube Data API v3, including both top-level comments and replies
3. **Preprocessing**: Comments were cleaned and normalized while preserving Korean linguistic characteristics
4. **Classification**: Each comment was processed through Google's Gemini 2.5 Flash-Lite model using carefully designed prompts to classify bot vs. human generation
5. **Quality Control**: A subset of classifications were manually reviewed to ensure label quality

#### Who are the source data producers?

The source data consists of comments from YouTube users who posted on popular Korean videos. The original content creators include a diverse range of Korean speakers, from casual viewers to active community members participating in discussions around trending Korean content.

### Annotations

#### Annotation process

Comments were automatically classified using Google's Gemini 2.5 Flash-Lite model. The classification process involved:
- Analyzing linguistic patterns, repetition, and semantic coherence
- Identifying typical bot behavior markers such as generic responses, promotional content, and unnatural language patterns
- Considering context relevance and engagement patterns
- Binary classification into bot (True) or human (False) categories

#### Who are the annotators?

Primary annotation was performed by Google's Gemini 2.5 Flash-Lite model, with quality validation through manual review of sample classifications.

## Personal and Sensitive Information

The dataset contains publicly available YouTube comments. While usernames and profile image URLs are included, all content was already public on the YouTube platform. No private or personal information beyond what users voluntarily shared in public comments is included. Users' real names, contact information, or other private details are not part of this dataset.

## Bias, Risks, and Limitations

**Limitations:**
- Classification accuracy is dependent on Gemini 2.5 Flash-Lite's performance, which may have inherent biases
- Limited to Korean language content and may not generalize to other languages
- Represents a snapshot of YouTube content from a specific time period
- Bot detection patterns may evolve, making the dataset less effective over time
- Source video selection may introduce topical or demographic biases

**Potential Biases:**
- May reflect biases present in the source videos' audiences
- Classification model may have systematic errors for certain types of content
- Temporal bias as bot techniques and patterns change over time

**Risks:**
- Potential misclassification of legitimate human comments as bot-generated
- May be used to develop more sophisticated bot systems
- Privacy concerns related to public comment analysis

### Recommendations

Users should:
- Validate model performance on their specific use case before deployment
- Be aware that bot detection is an evolving challenge with no perfect solutions
- Consider the temporal aspects of the data when applying to current content
- Implement additional validation steps for high-stakes applications
- Respect user privacy and platform terms of service when using insights from this research

## Citation

**BibTeX:**

```bibtex
@dataset{youtube_korean_bot_dataset,
  title={YouTube Korean Bot Comment Dataset},
  author={[MisileLab]},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/MisileLab/youtube-bot-comments]}
}
```

**APA:**

[MisileLab]. (2025). YouTube Korean Bot Comment Dataset. Hugging Face. https://huggingface.co/datasets/[MisileLab]/[youtube-bot-comments]

## Glossary

- **Bot Comment**: Automated or artificially generated comment, typically created by software rather than human users
- **YouTube Data API v3**: Google's official API for accessing YouTube data programmatically
- **Gemini 2.5 Flash-Lite**: Google's language model used for classification tasks
- **Top-level Comment**: Comments posted directly on a video (not replies to other comments)
- **Reply Comment**: Comments posted in response to other comments

## Dataset Card Authors

[MisileLab](https://huggingface.co/MisileLab)

## Dataset Card Contact

[email](mailto:misile@duck.com)