Datasets:

JohnnyGeorge
/

MasrawiQA

license: mit
task_categories:
  - question-answering
language:
  - arz
tags:
  - dialectal-arabic
  - cultural-commonsense
  - egyptian-arabic
size_categories:
  - n<1K

MasrawiQA

MasrawiQA is a culturally rooted, manually curated evaluation benchmark containing 129 question-answer pairs designed exclusively for the Cairene and Lower Egyptian Arabic dialect (arz_arab). Introduced as part of the MRL 2026 Shared Task, this dataset evaluates language models on authentic street slang, everyday idioms, and dense regional knowledge spanning Egyptian sports history, local cuisine, and urban pop culture.

Dataset Structure

The dataset consists of a single split containing the following features:

ID: Unique numeric identifier for each item.
Question: Casual Egyptian Arabic question text mapping real-world digital orthography.
Answer: Deterministic, short gold-standard solution string (1–3 words).
Automatic Translation: Literal English machine translation baseline.
Corrected Translation: Manually verified, context-accurate English translation capturing figurative intent.
Source URL: Verification URL linking to public archives confirming the underlying fact.

Usage

You can load this dataset directly using the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("JohnnyGeorge/MasrawiQA")
print(dataset['train'][0])