---
license: apache-2.0
language:
  - tr
  - ar
  - en
tags:
  - toxicity
  - multilingual
  - text-classification
size_categories:
  - 100K<n<1M
task_categories:
  - text-classification
---

# Multilingual Toxicity Dataset V3

124K balanced samples for binary toxicity classification across three languages.

## Author

- **[Görkem Yıldız](https://gorkemyildiz.com)**
- GitHub: [gorkem371](https://github.com/gorkem371)
- Website: [gorkemyildiz.com](https://gorkemyildiz.com)


## Dataset Details

| Split | Samples |
|-------|---------|
| Train | 105,945 |
| Valid | 18,684 |
| **Total** | **124,629** |

## Languages

- **Turkish (~34%)** — sourced primarily from Overfit-GM
- **Arabic (~32%)** — sourced primarily from arabic-hate-speech-superset
- **English (~34%)** — sourced from toxic_conversations_50k, hate_speech_offensive, and others

Label balance is approximately **50/50 toxic / not_toxic per language**.

## Format

JSONL files with the following fields:

| Field | Type | Description |
|-------|------|-------------|
| `text` | string | The input text |
| `label` | int | `1` = toxic, `0` = not toxic |
| `language` | string | Language code: `tr`, `ar`, or `en` |

## Sources

- **Overfit-GM** (Turkish)
- **arabic-hate-speech-superset** (Arabic)
- **toxic_conversations_50k** (English)
- **hate_speech_offensive** (English)
- **textdetox** (Multilingual)
- Targeted / hand-crafted samples for edge cases

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("gorkem371/toxicity-classifier-dataset-v3")
```

## License

Apache 2.0