Datasets:
Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,66 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
language:
|
| 4 |
+
- tr
|
| 5 |
+
- ar
|
| 6 |
+
- en
|
| 7 |
+
tags:
|
| 8 |
+
- toxicity
|
| 9 |
+
- multilingual
|
| 10 |
+
- text-classification
|
| 11 |
+
size_categories:
|
| 12 |
+
- 100K<n<1M
|
| 13 |
+
task_categories:
|
| 14 |
+
- text-classification
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# Multilingual Toxicity Dataset V3
|
| 18 |
+
|
| 19 |
+
124K balanced samples for binary toxicity classification across three languages.
|
| 20 |
+
|
| 21 |
+
## Dataset Details
|
| 22 |
+
|
| 23 |
+
| Split | Samples |
|
| 24 |
+
|-------|---------|
|
| 25 |
+
| Train | 105,945 |
|
| 26 |
+
| Valid | 18,684 |
|
| 27 |
+
| **Total** | **124,629** |
|
| 28 |
+
|
| 29 |
+
## Languages
|
| 30 |
+
|
| 31 |
+
- **Turkish (~34%)** — sourced primarily from Overfit-GM
|
| 32 |
+
- **Arabic (~32%)** — sourced primarily from arabic-hate-speech-superset
|
| 33 |
+
- **English (~34%)** — sourced from toxic_conversations_50k, hate_speech_offensive, and others
|
| 34 |
+
|
| 35 |
+
Label balance is approximately **50/50 toxic / not_toxic per language**.
|
| 36 |
+
|
| 37 |
+
## Format
|
| 38 |
+
|
| 39 |
+
JSONL files with the following fields:
|
| 40 |
+
|
| 41 |
+
| Field | Type | Description |
|
| 42 |
+
|-------|------|-------------|
|
| 43 |
+
| `text` | string | The input text |
|
| 44 |
+
| `label` | int | `1` = toxic, `0` = not toxic |
|
| 45 |
+
| `language` | string | Language code: `tr`, `ar`, or `en` |
|
| 46 |
+
|
| 47 |
+
## Sources
|
| 48 |
+
|
| 49 |
+
- **Overfit-GM** (Turkish)
|
| 50 |
+
- **arabic-hate-speech-superset** (Arabic)
|
| 51 |
+
- **toxic_conversations_50k** (English)
|
| 52 |
+
- **hate_speech_offensive** (English)
|
| 53 |
+
- **textdetox** (Multilingual)
|
| 54 |
+
- Targeted / hand-crafted samples for edge cases
|
| 55 |
+
|
| 56 |
+
## Usage
|
| 57 |
+
|
| 58 |
+
```python
|
| 59 |
+
from datasets import load_dataset
|
| 60 |
+
|
| 61 |
+
dataset = load_dataset("gorkem371/toxicity-classifier-dataset-v3")
|
| 62 |
+
```
|
| 63 |
+
|
| 64 |
+
## License
|
| 65 |
+
|
| 66 |
+
Apache 2.0
|