Tawkeed Arabic Benchmark is a curated dataset of 970 questions spanning multiple categories designed to rigorously evaluate Arabic and Saudi LLM capabilities across diverse tasks — from MMLU-style knowledge questions to dialect writing, diacritization, and Arabic grammar analysis.

Dataset Overview

Metric	Value
Total Questions	970
Languages	Arabic (MSA + Dialects)
Question Formats	MCQ, Generation
Scoring Methods	LLM-as-Judge + Manual metrics

Dataset Schema

Each record contains the following fields:

{
  "instruction": "the question in Arabic",
  "output": "the reference answer",
  "category": "Category Name",
  "subcategory": "Specific Subcategory",
  "format": "MCQ | Generation",
  "scoring_rules": ["SCORING_RULE_NAME"],
  "system_prompt": "",
  "choices": ["choice 1", "choice 2", "..."],
  "reference_lang": "ara"
}

Scoring Methods

LLM-as-Judge

AUTOMATED_LLM_AS_A_JUDGE_MCQ — Multiple choice evaluation
AUTOMATED_LLM_AS_A_JUDGE_GENERATION — Open-ended generation quality
AUTOMATED_LLM_AS_A_JUDGE_WRITING_DIALECT — Dialect writing accuracy
AUTOMATED_LLM_AS_A_JUDGE_REASONING — Reasoning and math evaluation
AUTOMATED_LLM_AS_A_JUDGE_GRAMMAR_IRAB — Arabic grammar analysis

Manual Scoring

MANUAL_ROUGE_SCORE — ROUGE-L F1 for summarization
MANUAL_METEOR_SCORE — METEOR for translation
MANUAL_WORDS_INTERSECTION — Word overlap for entity extraction
MANUAL_DIACRITIZATION — Character-level diacritization accuracy
MANUAL_MIN_DISTANCE — Levenshtein-based similarity
MANUAL_IS_VALID_JSON — JSON validity for function calling
MANUAL_IFEVAL_1 / MANUAL_IFEVAL_2 — Instruction following
MANUAL_STRUCTURING_1 — Structural formatting evaluation

Usage

from datasets import load_dataset

# Requires accepting terms of use first
dataset = load_dataset("tawkeed-sa/tawkeed-arabic-benchmark", split="test")

print(f"Total questions: {len(dataset)}")
print(dataset[0])

Prerequisites

Running the benchmark evaluation requires:

Requirement	Description
OpenAI API Key	The scoring judge uses GPT-5.2 via the OpenAI API to evaluate model responses. Set `OPENAI_API_KEY` in your environment.
Model API Endpoint	An OpenAI-compatible chat completions endpoint for the model you want to evaluate.

export OPENAI_API_KEY="sk-..."   # Required for GPT-5.2 judge scoring

Evaluation

Use the Tawkeed Benchmark Runner to evaluate your model:

python benchmark.py

Citation

@dataset{tawkeed_arabic_benchmark_2026,
  title={Tawkeed Arabic Benchmark},
  author={Tawkeed Team},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/tawkeed-sa/tawkeed-arabic-benchmark}
}

License

This dataset is released under the Tawkeed Benchmark License. See LICENSE.md for full terms.

By accessing this dataset, you agree to:

Use the dataset for legitimate evaluation and research purposes
Not redistribute the dataset without permission
Cite Tawkeed Arabic Benchmark in any publications using this data
Not use the dataset to train models that generate harmful content
Comply with all applicable laws and regulations

Contact

For questions, feedback, or access issues, contact the Tawkeed team.

Downloads last month: 10

Total file size:

1.25 MB

Category	Count
MMLU	331
General Knowledge	103
Reading Comprehension	77
Reasoning & Math	73
RAG QA	71
Sentiment Analysis	59
Arabic Language & Grammar	47
Translation (incl Dialects)	36
Hallucination	33
Trust & Safety	30
Coding	23
Writing (incl Dialects)	22
Diacritization	12
Dialect Detection	11
Summarization	8
Instruction Following	7
Transliteration	6
Paraphrasing	6
Entity Extraction	5
Long Context	4
Function Calling	3
Structuring	3