| --- |
| language: |
| - ar |
| license: apache-2.0 |
| task_categories: |
| - text-classification |
| task_ids: |
| - sentiment-classification |
| - hate-speech-detection |
| tags: |
| - arabic |
| - toxicity |
| - hate-speech |
| - cyberbullying |
| - content-moderation |
| - offensive-language |
| - profanity |
| - multi-dialect |
| - arabic-nlp |
| pretty_name: "Hayā — Arabic Toxic Content Dataset" |
| size_categories: |
| - 100K<n<1M |
| source_datasets: |
| - original |
| - L-HSAB |
| - T-HSAB |
| - OSACT4 |
| - OSACT5 |
| - MPOLD |
| - Let-Mi |
| - ArMIS |
| - ArMI |
| - ADHAR |
| - LREC |
| dataset_info: |
| features: |
| - name: text |
| dtype: string |
| - name: label |
| dtype: int64 |
| - name: source |
| dtype: string |
| - name: dialect |
| dtype: string |
| splits: |
| - name: train |
| num_examples: 798071 |
| - name: validation |
| num_examples: 99759 |
| - name: test |
| num_examples: 99759 |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: train.parquet |
| - split: validation |
| path: val.parquet |
| - split: test |
| path: test.parquet |
| --- |
| |
| # Hayā (حياء) — Arabic Toxic Content Dataset 🛡️ |
|
|
| [](https://huggingface.co/youssefreda9/HAYAA) |
| [](https://github.com/youssefreda10/HAYAA) |
|
|
| ## Dataset Description |
|
|
| **Hayā** is a large-scale, curated Arabic binary toxicity dataset built for training and evaluating Arabic content-moderation models. It was assembled by reviewing, collecting, standardizing, and merging **51 public Arabic hate-speech and abuse datasets** into a single clean corpus of **nearly 1 million examples**. |
|
|
| This is the dataset used to train the [Hayā model](https://huggingface.co/youssefreda9/HAYAA), a fine-tuned **UBC-NLP/MARBERTv2** classifier that achieved **97.84% accuracy** on the held-out test set. |
|
|
| ### Key Features |
|
|
| - **~997K examples** of Arabic text (comments, posts, tweets) labeled as **Safe (0)** or **Toxic (1)** |
| - **Every major Arabic dialect**: Egyptian, Levantine, Gulf, Maghrebi, Iraqi, Sudanese, and MSA |
| - **Multi-platform**: Twitter, Facebook, YouTube, news comments, and more |
| - **Deduplicated**: Zero duplicate texts across the entire corpus |
| - **Conflict-resolved**: Overlapping texts from different sources resolved via majority vote |
| - **Leak-free splits**: Stratified 80/10/10 split with verified zero overlap between train, validation, and test |
|
|
| --- |
|
|
| ## Dataset Summary |
|
|
| | Property | Value | |
| |----------|-------| |
| | **Language** | Arabic (all major dialects) | |
| | **Task** | Binary text classification | |
| | **Labels** | `0` = Safe, `1` = Toxic | |
| | **Total Size** | 997,589 examples | |
| | **Train** | 798,071 (80%) | |
| | **Validation** | 99,759 (10%) | |
| | **Test** | 99,759 (10%) | |
| | **Sources** | 51 public datasets merged & standardized | |
| | **Format** | Parquet | |
|
|
| --- |
|
|
| ## What "Toxic" Covers |
|
|
| The **Toxic** label encompasses: |
|
|
| - 🤬 **Profanity & offensive language** |
| - 🔥 **Hate speech** |
| - 😡 **Insults, harassment & cyberbullying** |
| - 🚫 **Racism & religious hate** |
| - ⚠️ **Sexism & sexually explicit language** |
| - 🔤 **Morphologically complex and context-dependent expressions** |
| - 🕵️ **Intentional typos / obfuscation attempts** |
|
|
| --- |
|
|
| ## Dialects Covered |
|
|
| | Dialect Group | Varieties | |
| |---------------|-----------| |
| | **Egyptian** | Egyptian Arabic | |
| | **Levantine** | Syrian, Lebanese, Jordanian, Palestinian | |
| | **Gulf** | Saudi, Emirati, Kuwaiti, Bahraini, Omani, Qatari | |
| | **Maghrebi** | Moroccan, Algerian, Tunisian, Libyan | |
| | **Other** | Iraqi, Sudanese, Yemeni | |
| | **MSA** | Modern Standard Arabic | |
|
|
| --- |
|
|
| ## Data Fields |
|
|
| | Field | Type | Description | |
| |-------|------|-------------| |
| | `text` | `string` | The Arabic text (comment, tweet, or post) | |
| | `label` | `int` | `0` = Safe, `1` = Toxic | |
| | `source` | `string` | Origin dataset(s), pipe-separated if from multiple sources | |
| | `dialect` | `string` | Dialect(s), pipe-separated if from multiple sources | |
|
|
| --- |
|
|
| ## Data Splits |
|
|
| | Split | Examples | Purpose | |
| |-------|----------|---------| |
| | `train` | ~798K | Model training | |
| | `validation` | ~100K | Hyperparameter tuning & early stopping | |
| | `test` | ~100K | Final held-out evaluation | |
|
|
| All splits are **stratified by label** and verified for **zero text overlap** (no data leakage). |
|
|
| --- |
|
|
| ## Usage |
|
|
| ```python |
| from datasets import load_dataset |
| |
| dataset = load_dataset("youssefreda9/HAYAA") |
| |
| # Access splits |
| train = dataset["train"] |
| val = dataset["validation"] |
| test = dataset["test"] |
| |
| # Example |
| print(train[0]) |
| # {'text': '...', 'label': 0, 'source': 'OSACT4', 'dialect': 'MSA'} |
| ``` |
|
|
| --- |
|
|
| ## Curation Process |
|
|
| ### 1. Collection |
| Reviewed and collected **100+ public Arabic hate-speech and abuse datasets** from published research, shared tasks, and public repositories. |
|
|
| ### 2. Standardization |
| Every source used a different labeling scheme (3-class, 5-class, multi-label, etc.). All were standardized into a single **binary** scheme: |
| - **Safe (0)**: Non-toxic, normal text |
| - **Toxic (1)**: Offensive, hateful, abusive, or harmful |
|
|
| ### 3. Cleaning & Quality Control |
| - **Removed problematic datasets**: Sources with noisy labels, wrong task definitions, or exact duplicates of other sources were excluded |
| - **Fixed known mislabels**: Manual keyword-based correction of known systematic errors (e.g., SaudiCodeMixing safe-labeled profanity) |
| - **Filtered noise**: Removed long, machine-translated safe texts that didn't match real Arabic usage patterns |
|
|
| ### 4. Conflict Resolution |
| Texts appearing in multiple sources with conflicting labels were resolved via **majority vote**. |
|
|
| ### 5. Deduplication |
| Full text-level deduplication with dialect and source metadata merged (pipe-separated) for multi-source texts. |
|
|
| ### 6. Splitting |
| **Stratified 80/10/10 split** (train/val/test) with seed=42, verified for zero leakage across all split pairs. |
|
|
| --- |
|
|
| ## Source Datasets |
|
|
| This corpus merges **51 standardized public Arabic datasets**, including: |
|
|
| | Dataset | Dialect | Size | Platform | |
| |---------|---------|------|----------| |
| | L-HSAB | Levantine | ~5.8K | Twitter | |
| | T-HSAB | Tunisian | ~6K | Facebook/YouTube | |
| | OSACT4 | MSA + mixed | ~10K | Twitter | |
| | OSACT5 | MSA + mixed | ~12.7K | Twitter | |
| | MPOLD | Multi-dialect | ~4K | Multi-platform | |
| | Let-Mi | Levantine | ~6.6K | Twitter | |
| | ArMI | MSA, Egyptian, Gulf, Levantine | ~9.8K | Twitter | |
| | ADHAR | Multi-dialect | ~4.3K | Twitter | |
| | LREC | MSA + dialects | ~15.9K | Twitter | |
| | *... and 42 more* | | | | |
|
|
| See the full [Dataset Registry](https://github.com/youssefreda10/HAYAA/tree/main/data) for detailed per-source documentation. |
|
|
| --- |
|
|
| ## Model Trained on This Dataset |
|
|
| The [Hayā model](https://huggingface.co/youssefreda9/HAYAA) — fine-tuned **UBC-NLP/MARBERTv2** — achieves: |
|
|
| | Metric | Score | |
| |--------|-------| |
| | Accuracy | **97.84%** | |
| | F1 (Toxic class) | **94.12%** | |
| | F1 (Safe class) | 98.45% | |
|
|
| > **Note:** Manual error analysis showed the model frequently *outperformed the original human annotations* — many counted "errors" were actually mislabels in the source data. |
|
|
| --- |
|
|
| ## Intended Use |
|
|
| - ✅ Training and evaluating Arabic toxicity / hate-speech classifiers |
| - ✅ Benchmarking Arabic NLP models on content moderation |
| - ✅ Research on multi-dialect Arabic text classification |
| - ✅ Building content-moderation pipelines for Arabic platforms |
|
|
| ## Out-of-Scope Use |
|
|
| - ❌ Generating toxic or hateful content |
| - ❌ Surveillance or profiling of individuals |
| - ❌ Deployed decisions without human review |
|
|
| --- |
|
|
| ## Limitations & Biases |
|
|
| - **Platform bias**: The majority of the data comes from Twitter; other platforms (forums, messaging apps) are underrepresented. |
| - **Annotation noise**: Despite extensive cleaning, some label noise from original sources may remain. |
| - **Dialect imbalance**: Some dialects (e.g., Egyptian, MSA) are better represented than others (e.g., Sudanese, Yemeni). |
| - **Temporal bias**: Most data reflects Arabic social media discourse from 2018–2024; slang and evasion tactics evolve constantly. |
|
|
| --- |
|
|
| ## Citation |
|
|
| If you use this dataset, please cite: |
|
|
| ```bibtex |
| @misc{hayaa2026, |
| title={Hayā: A Large-Scale Multi-Dialect Arabic Toxicity Dataset}, |
| author={Youssef Reda}, |
| year={2026}, |
| url={https://huggingface.co/datasets/youssefreda9/HAYAA}, |
| } |
| ``` |
|
|
| --- |
|
|
| ## License |
|
|
| This dataset is released under the [Apache 2.0 License](https://www.apache.org/licenses/LICENSE-2.0). |
|
|
| --- |
|
|
| ## Contact |
|
|
| - **GitHub**: [youssefreda10/HAYAA](https://github.com/youssefreda10/HAYAA) |
| - **Model**: [youssefreda9/HAYAA](https://huggingface.co/youssefreda9/HAYAA) |
|
|