sbunlp/FarsEval

Benchmark Description

FarsEval is an ongoing initiative for Persian language evaluation, aimed at assessing the capabilities of large language models (LLMs) in Persian while accounting for the linguistic, cultural, and local characteristics of the language, particularly in the Iranian context.

As part of this broader project, FarsEval-PKBETS (Persian Knowledge, Bias, Ethics, Toxicity and Skills) is an open-source benchmark comprising 4,000 question–answer pairs in diverse formats, including descriptive, short-answer, and multiple-choice questions. It covers a wide range of domains and evaluation aspects, including medicine, law, religion, Persian language, encyclopedic and social knowledge, human preferences, NLP tasks, ethics and bias, toxicity, respect for others’ rights, and text generation. All samples are carefully curated and expert-validated to ensure correctness and to reflect the linguistic, cultural, and local characteristics of Persian and the Iranian context.

In addition, FarsEval-Boom extends the FarsEval project with 300 records covering topics such as culinary, Iranian tourism and traditions, politics, history, and summarization. The name Boom (in Persian:بوم) is a Persian word meaning “land,” reflecting the benchmark’s emphasis on Iran’s cultural heritage and local knowledge.

Acknowledgement

We would like to sincerely thank the National Artificial Intelligence Organization (NAIO) for providing the funds necessary to support this study. We also gratefully acknowledge the Vice-Presidency for Science, Technology and Knowledge-Based Economy and the Artificial Intelligence Technology Development and Application Headquarters for their support in making the dataset openly available.

Citation Information

If you use FarsEval in your research or applications, please cite the following paper(s).

@misc{shamsfard2025farsevalpkbetsnewdiversebenchmark,
      title={FarsEval-PKBETS: A new diverse benchmark for evaluating Persian large language models}, 
      author={Mehrnoush Shamsfard and Zahra Saaberi and Mostafa Karimi manesh and Seyed Mohammad Hossein Hashemi and Zahra Vatankhah and Motahareh Ramezani and Niki Pourazin and Tara Zare and Maryam Azimi and Sarina Chitsaz and Sama Khoraminejad and Morteza Mahdavi Mortazavi and Mohammad Mahdi Chizari and Sahar Maleki and Seyed Soroush Majd and Mostafa Masumi and Sayed Ali Musavi Khoeini and Amir Mohseni and Sogol Alipour},
      year={2025},
      eprint={2504.14690},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2504.14690}, 
}

@inproceedings{hosseinbeigi-etal-2025-advancing,
    title = "Advancing {P}ersian {LLM} Evaluation",
    author = "Hosseinbeigi, Sara Bourbour  and
      Rohani, Behnam  and
      Masoudi, Mostafa  and
      Shamsfard, Mehrnoush  and
      Saaberi, Zahra  and
      Manesh, Mostafa Karimi  and
      Abbasi, Mohammad Amin",
    editor = "Chiruzzo, Luis  and
      Ritter, Alan  and
      Wang, Lu",
    booktitle = "Findings of the Association for Computational Linguistics: NAACL 2025",
    month = apr,
    year = "2025",
    address = "Albuquerque, New Mexico",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.findings-naacl.147/",
    doi = "10.18653/v1/2025.findings-naacl.147",
    pages = "2711--2727",
    ISBN = "979-8-89176-195-7",
    abstract = "Evaluation of large language models (LLMs) in low-resource languages like Persian has received less attention than in high-resource languages like English. Existing evaluation approaches for Persian LLMs generally lack comprehensive frameworks, limiting their ability to assess models' performance over a wide range of tasks requiring considerable cultural and contextual knowledge, as well as a deeper understanding of Persian literature and style. This paper first aims to fill this gap by providing two new benchmarks, PeKA and PK-BETS, on topics such as history, literature, and cultural knowledge, as well as challenging the present state-of-the-art models' abilities in a variety of Persian language comprehension tasks. These datasets are meant to reduce data contamination while providing an accurate assessment of Persian LLMs. The second aim of this paper is the general evaluation of LLMs across the current Persian benchmarks to provide a comprehensive performance overview. By offering a structured evaluation methodology, we hope to promote the examination of LLMs in the Persian language."
}