README.md

metadata

dataset_info:
  features:
    - name: text
      dtype: string
    - name: prediction
      struct:
        - name: prediction_confidence
          dtype: float64
        - name: prediction_label
          dtype: string
  splits:
    - name: train
      num_bytes: 5595607628.85211
      num_examples: 21420806
  download_size: 2351016812
  dataset_size: 5595607628.85211
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Moroccan Darija Dataset (Filtered from FineWeb2)

This dataset contains Moroccan Darija samples extracted from the FineWeb2 dataset. The extraction was performed using a custom model trained to classify Arabic dialects, including Moroccan Darija (here we used the version model_binary_v3_1fpr.bin). Here, we kept samples with a high confidence score (above 0.9) for Moroccan Darija. This dataset aims to advance research and development in Moroccan Darija NLP tasks.

---

Dataset Description

Moroccan Darija, a widely spoken Arabic dialect in Morocco, is underrepresented in NLP resources. This dataset fills that gap by filtering FineWeb2 using an advanced classifier designed to accurately identify Moroccan Darija text. The resulting dataset is a valuable resource for tasks such as:

• Language modeling
• Sentiment analysis
• Machine translation
• Dialectal classification

---

Extraction Methodology

1. Base Dataset: FineWeb2, a large-scale multilingual web dataset.
2. First extraction using GlotLID: A version with extraction using GlotLID.
3. SfaIA Model: A fasttext model trained to identify Arabic dialects, including Moroccan Darija with better performances than GlotLID.
4. Pipeline:
   • Text samples from the dataset were passed through the SfaIA classifier.
   • Only samples with a high confidence score (above 0.9) for Moroccan Darija were retained.

---

Dataset Structure

• text: The raw text sample in Moroccan Darija.
• prediction:
  • prediction_confidence: Model confidence score for each sample.
  • prediction_label: Model predicted label.

Example entry:

{
  "text": "السلام عليكم، كيف دايرين؟",
  "prediction": {
      'prediction_confidence': 0.952466607093811, 
      'prediction_label': 'Morocco'
   }
}