EtMmohammedHafsati/darija-lid-benchmark

This repository contains the final merged Darija language identification dataset prepared from multiple public sources.

What Was Done

• Downloaded the configured source datasets.
• Flattened each source into one shared schema: text, dialect, src, has_latin_script, label.
• Mapped labels to three classes:
• 0: arabic_moroccan
• 1: arabizi_moroccan
• 2: other
• Added English translation rows when they were explicitly available in source datasets.
• Kept Moroccan Darija in two buckets:
  • Arabic-script Moroccan
  • Arabizi / Latin-script Moroccan
• Removed repeated normalized lines so the same text appears only once in the final dataset.
• Preserved provenance in the src column. If a row came from multiple sources, src stores them joined by |.

Dataset Schema

• text: normalized text
• dialect: normalized dialect name
• src: source dataset identifier(s)
• has_latin_script: whether the row contains at least one Latin character
• label: numeric class ID

Label Counts

• arabic_moroccan (0): 213,712
• arabizi_moroccan (1): 12,468
• other (2): 245,222

Build Summary

• Final row count: 471,402
• Input rows before deduplication: 533,380
• Output rows after deduplication: 471,402
• Duplicate rows removed: 61,978
• Duplicate groups with label or dialect conflicts: 139
• Raw rows excluded because they belonged to conflict groups: 527

Example Conflict Cases

Sampled from the middle of the conflict list:

• حلو، شكرا لك. (labels=[2], dialects=['saudi_arabic_hijazi_dialect', 'saudi_arabic_southern_dialect'], rows=2)
• دقيقة (labels=[0, 2], dialects=['arabic_moroccan', 'jordan'], rows=5)
• رائع (labels=[0, 2], dialects=['arabic_moroccan', 'msa'], rows=4)
• ربع ساعة (labels=[0, 2], dialects=['arabic_moroccan', 'msa'], rows=4)
• سلام عليكم. (labels=[2], dialects=['egyptian_arabic_cairene_dialect', 'levant'], rows=3)

Source Datasets

• atlasia/Darija_LID_Anootation_10k
• atlasia/DODa-audio-dataset
• atlasia/Darija-LID
• UBC-NLP/alexandria
• atlasia/levantine_dialects

Notes

• This is a text classification dataset for Darija language identification.
• The other class includes non-Moroccan dialects, English translation rows, and the remaining non-target content.
• The Hugging Face data viewer is enabled through the test.csv file declared above.

Visual Summary

Final Dataset Dashboard