Omartificial-Intelligence-Space/Arabic-Cohere-include-base-44-mmlu-style

The Refined Arabic Cohere INCLUDE Base 44 Dataset as MMLU-Style

Dataset Summary

INCLUDE is a comprehensive knowledge- and reasoning-centric benchmark spanning 44 languages that evaluates multilingual LLMs in the actual linguistic environments where they are deployed. The original dataset contains 22,637 4-option multiple-choice questions (MCQs) extracted from academic and professional exams, covering 57 topics, including regional knowledge.

When we reviewed the Arabic portion of this dataset, we discovered numerous issues:

🔸 Spelling Errors: 70% of questions and options contained significant spelling mistakes, making them difficult to read and evaluate.

🔸 Incorrect Answers: Some questions were paired with incorrect or misleading answers espically in Islamic Question which is very criticl

🔸 Misleading Information: Some questions conveyed incorrect or ambiguous meanings.

🔸 General Poor Quality: Around 80% of the dataset required revisions due to foundational errors.

Given the extent of these problems, we decided to use Argilla to meticulously review and refine the dataset. You can access the review process on the following Hugging Face Space

This refined version addresses foundational errors and enhances the overall quality of the dataset.

Foundational Issues Identified

During the review, we identified the following key issues in the original Arabic dataset:

◾️ Spelling Errors:

◾️ Example: Words were written incorrectly, often to the point where they were unreadable, even by native speakers.

◾️ Incorrect Answers:

◾️ Example: A question asking about a good trait of Arabs before Islam had the worst possible trait listed as the correct answer.

◾️ Ambiguity in Choices:

◾️ Example: Some questions had identical or nonsensical options, making it impossible to determine the correct answer logically.

◾️ Misleading Questions:

◾️ Example: Questions framed in ways that conveyed misleading or outright wrong information about Islamic teachings or Arabic culture.

The Refined Version

This refined version of the Arabic portion of the INCLUDE dataset resolves these issues. Key improvements include:

🔸 Corrected Spelling: All spelling mistakes have been fixed, ensuring readability and clarity.

🔸 Accurate Answers: Incorrect answers were replaced with the correct ones after thorough verification.

🔸 Clearer Questions and Choices: Questions and options have been rewritten to eliminate ambiguity and improve logical coherence.

🔸 Misleading Information Removed: All misleading or factually incorrect content has been corrected.

This refined Arabic dataset is a significant improvement over the original version, ensuring that it meets the standards required for evaluating multilingual large language models (LLMs). By addressing spelling errors, correcting factual inaccuracies, and eliminating ambiguity, this dataset now provides a reliable benchmark for Arabic LLM evaluation.