TajikMMLU is the largest Multiple Choice Question Answering (MCQA) dataset in the Tajik language, containing 100,215 questions across 599 subjects spanning diverse knowledge domains. The dataset is designed for training and evaluating MCQA models for low-resource languages, with particular focus on Central Asian linguistic diversity.

Curated by: TajikNLPWorld
Language(s) (NLP): Tajik (tg)
License: other – see Licensing & Legal Notice below.

Licensing & Legal Notice

This dataset follows the practice established by large web‑crawled corpora such as HPLT and OSCAR:

Original question texts (questions, answer options) remain the property of their respective authors and educators. They are not owned by the TajikNLPWorld team and are not covered by the CC0 1.0 dedication.
The structured compilation, metadata, annotations (correct answer labels), and packaging created by TajikNLPWorld are released under the CC0 1.0 Universal Public Domain Dedication.
Users are solely responsible for ensuring their use of the underlying content complies with applicable copyright law. For commercial use of verbatim excerpts, permission from the original copyright holders may be required.
A notice‑and‑takedown procedure is in place: rights holders can request removal of specific content by contacting the dataset maintainers (see Dataset Card Contact). We commit to responding within 14 business days and removing disputed content in the next release.

Dataset Sources

Repository: https://huggingface.co/datasets/TajikNLPWorld/TajikMMLU

Uses

Direct Use

The dataset is intended for:

Training and evaluating multiple-choice question answering models for the Tajik language.
Benchmarking low-resource language models on knowledge-intensive tasks.
Research on cross-lingual transfer and domain adaptation for Central Asian languages.
Educational applications such as automated quiz generation or assessment tools.

Out-of-Scope Use

The dataset should not be used:

As a source of truth for critical decisions (e.g., medical, legal, or safety-critical systems).
For any purpose that violates the original copyright of the question texts (see Licensing & Legal Notice).
To discriminate against or misrepresent any group of people.

Dataset Structure

The dataset consists of a single split (train) with 100,215 examples. Each example is a multiple-choice question with exactly four answer options and one correct answer.

Data Fields

Field	Type	Description
`question_id`	string	Unique question identifier
`subject`	string	Academic subject name (in Tajik)
`category`	string	Knowledge domain category (one of 9 categories)
`question_text`	string	The question text (in Tajik)
`option_1`	string	Answer option A
`option_2`	string	Answer option B
`option_3`	string	Answer option C
`option_4`	string	Answer option D
`correct_answer`	string	The correct answer text
`correct_option`	string	Position of correct answer (`option_1`, `option_2`, `option_3`, or `option_4`)
`question_char_count`	int32	Character count of the question
`total_options_char_count`	int32	Total character count of all four options
`question_token_count`	int32	Word token count of the question

Example Record

Field	Value
`question_id`	`1`
`subject`	`Методикаи таълими фанҳои педагоги`
`category`	`Humanities`
`question_text`	`Методика таълимот чиро мефаҳмонад?`
`option_1`	`методикаи таълимот усулҳои дарсдиҳиро мефаҳмонад`
`option_2`	`назарияи тарбияро ифода мекунад`
`option_3`	`таърихи педагогикаро мефаҳмона`
`option_4`	`маҳорати педагогиро мефаҳмона`
`correct_answer`	`методика таълимот усулҳои дарсдиҳиро мефаҳмонад`
`correct_option`	`option_1`
`question_char_count`	`31`
`total_options_char_count`	`90`
`question_token_count`	`4`

Knowledge Categories

Category	Questions	Percentage
Humanities	33,490	33.4%
Other	22,544	22.5%
Natural Sciences	14,844	14.8%
Economics & Business	13,882	13.9%
Engineering & Technology	11,122	11.1%
Sports & Tourism	2,749	2.7%
Medicine & Healthcare	932	0.9%
Agriculture	491	0.5%
Military Science	161	0.2%

Dataset Creation

Curation Rationale

Tajik, a low-resource language of the Persian family, lacks large-scale benchmarks for knowledge-intensive tasks. TajikMMLU was created to fill this gap by providing a diverse, high-quality MCQA dataset that can serve as a foundation for training and evaluating models in Tajik, as well as for studying cross-lingual transfer from higher-resource languages.

Source Data

Data Collection and Processing

The dataset was collected by the authors during their work at the Tajik National University from 2012 to 2016, and subsequently expanded through collaborations with colleagues from the following institutions:

Tajik National University
Tajik State Pedagogical University named after Sadriddin Aini
Bokhtar State University named after Nosir Khusrav
Other higher education institutions of Tajikistan

The collection process continued over the period 2012–2026. Questions were compiled from various academic sources, examination materials, and educational resources in the Tajik language. Data processing included:

Filtering out duplicates and low-quality questions.
Standardizing the format (four options, one correct answer).
Assigning each question to a subject and a high-level knowledge category.
Computing character and token counts for analysis.

No automated translation was used; all questions are originally in Tajik.

Who are the source data producers?

The source data producers are educators, lecturers, and content creators from the Tajik universities listed above. The questions reflect the curricula and examination standards of Tajik higher education institutions during the collection period.

Annotations

Annotation process

Annotations (correct answers) were provided by the same educators who authored or curated the questions. No separate annotation campaign was conducted.

Who are the annotators?

The annotators are the same individuals who created the questions — subject matter experts from Tajik universities.

Personal and Sensitive Information

The dataset consists of general knowledge and educational questions. It does not contain personal data, addresses, names, or any information that could be used to identify individuals. No anonymization was necessary.

Bias, Risks, and Limitations

Over‑representation of certain subjects or domains (e.g., Humanities).
Under‑representation of rapidly evolving fields.
Possible factual inaccuracies or outdated information due to the long collection period.
Limited to the Tajik language; may not generalize to other Persian dialects.

Recommendations

Be aware of the domain distribution when evaluating models.
Avoid using the dataset for high-stakes decisions without additional validation.
Respect the copyright of individual question texts (see Licensing & Legal Notice).
Consider combining with other resources for a more balanced evaluation.

Citation

BibTeX (dataset):

@misc{tajikmmlu2026,
  author = {TajikNLPWorld},
  title = {TajikMMLU: A Large-Scale Multiple Choice Question Answering Dataset for Tajik},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/TajikNLPWorld/TajikMMLU}
}

APA (dataset): TajikNLPWorld. (2026). TajikMMLU: A Large-Scale Multiple Choice Question Answering Dataset for Tajik [Data set]. Hugging Face. https://huggingface.co/datasets/TajikNLPWorld/TajikMMLU

Glossary

MCQA – Multiple Choice Question Answering: a task where a model selects the correct answer from several options.
MMLU – Massive Multitask Language Understanding: a benchmark style covering many subjects and domains.
Low-resource language – a language with limited digital resources for NLP.

More Information

For additional details, updates, or to report issues, please visit the Hugging Face dataset repository.

Dataset Card Authors

TajikNLPWorld (community)

Dataset Card Contact

For questions, takedown requests, or collaboration, please open an issue on the Hugging Face repository:
https://huggingface.co/datasets/TajikNLPWorld/TajikMMLU/discussions
📧 Direct email: cool.araby@gmail.com

Downloads last month: 14

Total file size:

15.3 MB