TAAG — Teacher-Authored Arabic Grammar

TAAG is a small evaluation benchmark of 171 hand-written multiple-choice questions in Modern Standard Arabic, designed to measure how well a language model handles Arabic as a language, including its syntax, morphology, rhetoric, and lexical semantics.

Every question was written from scratch by a certified Arabic-language teacher and follows material inspired by the Arabic curriculum used in Lebanese schools.

TAAG was developed as part of the evaluation framework for NOHA, a diacritic-aware Arabic tokenizer published in Transactions on Machine Learning Research (TMLR) in 2026.

Associated publication
Ali Issa, Tina Yaacoub, and Patrick Abi Salloum.
NOHA: A Diacritic-Aware Rule-Based Morphophonological Tokenizer for Arabic.
Transactions on Machine Learning Research (TMLR), 2026.
https://openreview.net/forum?id=Vyc4Ormkqp

What is in it

Every item contains a question with three or four answer options and exactly one correct answer.

The questions cover:

• النحو (syntax) — case marking (الإعراب), sentence structure, grammatical functions, particles, and agreement.
• الصرف (morphology) — word patterns (الأوزان), derivation, root-and-pattern analysis, and verb forms.
• البلاغة (rhetoric) — figures of speech, simile, metaphor, and the interpretation of stylistic constructions.
• Lexical semantics — including synonyms, antonyms, and odd-one-out questions.

Provenance

The questions were written from scratch by a certified Arabic-language teacher with 15 years of classroom experience.

They were:

• not scraped from online sources,
• not translated from another language,
• not generated by a language model.

Their content and difficulty are inspired by the official Arabic-language textbooks used in Lebanese schools for the intermediate stage (المرحلة المتوسّطة), corresponding to grades 7, 8, and 9.

The questions are original compositions informed by that curriculum and are not copies of textbook exercises.

Dataset structure

TAAG contains a single test split of 171 questions.

There is no training split because the dataset is intended exclusively for evaluation.

Field	Type	Description
id	string	Stable item identifier, e.g. TAAG_000. IDs are zero-padded but are not necessarily contiguous.
question	string	The question stem in Arabic.
choices	list[string]	Answer options in their original order. Each item contains 3 or 4 options.
answer	string	Correct option represented as A, B, C, or D.
answer_index	int32	Zero-based index of the correct option in choices.
answer_text	string	Text of the correct answer.

Example

{
    "id": "TAAG_000",
    "question": "مرادف كلمة ممتنّة هو",
    "choices": ["راضية", "شاكرة", "صارحة"],
    "answer": "B",
    "answer_index": 1,
    "answer_text": "شاكرة",
}

Composition

	Count
Total questions	171
Questions with 4 options	148
Questions with 3 options	23

The correct-answer distribution is:

Answer	Count
A	28
B	66
C	56
D	21

The options were not shuffled to artificially balance answer positions.

Researchers evaluating models by answer letter rather than by option content should therefore account for this distribution. For example, a system with a systematic preference for option B may obtain an inflated score.

Usage

The dataset can be loaded directly using the Hugging Face datasets library:

from datasets import load_dataset

ds = load_dataset("ali-issa/TAAG", split="test")

for item in ds:
    options = "\n".join(
        f"{chr(65 + i)}) {choice}"
        for i, choice in enumerate(item["choices"])
    )

    prompt = f"{item['question']}\n{options}\nالجواب:"

    # Score the model's response against item["answer"]

Arabic text and diacritization

The dataset is distributed in undiacritized Modern Standard Arabic, reflecting the way Arabic is normally written in most contemporary text.

If an evaluation setup requires fully or partially diacritized text (تشكيل), researchers may apply a diacritization system of their choice.

TAAG is intentionally released in undiacritized form so that the choice of diacritization procedure is not embedded in the benchmark itself.

Use in NOHA

TAAG was introduced as part of the evaluation framework used in:

NOHA: A Diacritic-Aware Rule-Based Morphophonological Tokenizer for Arabic

The benchmark was used to assess Arabic linguistic capabilities that are difficult to capture using general-knowledge evaluation alone, particularly:

• syntax,
• morphology,
• rhetoric,
• lexical knowledge.

The associated paper evaluates NOHA against standard BPE tokenization under matched model architecture and compute conditions.

Paper:

https://openreview.net/forum?id=Vyc4Ormkqp

Citation

If you use TAAG in academic work, please cite the associated TMLR publication:

@article{
issa2026noha,
title={{NOHA}: A Diacritic-Aware Rule-Based Morphophonological Tokenizer for Arabic},
author={Ali Issa and Tina Yaacoub and Patrick Abi Salloum},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2026},
url={https://openreview.net/forum?id=Vyc4Ormkqp},
note={}
}

When referring specifically to the benchmark, please use its full name:

TAAG: Teacher-Authored Arabic Grammar

and include a link to this Hugging Face dataset repository when appropriate.

License

TAAG is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

You may use, share, and adapt the dataset, including for commercial purposes, provided that appropriate attribution is given.

License information:

https://creativecommons.org/licenses/by/4.0/

Associated publication

Ali Issa, Tina Yaacoub, and Patrick Abi Salloum.
NOHA: A Diacritic-Aware Rule-Based Morphophonological Tokenizer for Arabic.
Transactions on Machine Learning Research (TMLR), 2026.

Paper:

https://openreview.net/forum?id=Vyc4Ormkqp

Authors

Ali Issa
Tina Yaacoub
Patrick Abi Salloum

Université Saint-Joseph de Beyrouth

Contact

Questions, corrections, and reports concerning individual questions are welcome.

Please open a discussion on this Hugging Face repository.