metadata
configs:
- config_name: ar
data_files:
- split: top
path: ar/top.parquet
- split: high
path: ar/high.parquet
- split: medium
path: ar/medium.parquet
- split: low
path: ar/low.parquet
- config_name: bn
data_files:
- split: top
path: bn/top.parquet
- split: high
path: bn/high.parquet
- split: medium
path: bn/medium.parquet
- split: low
path: bn/low.parquet
- config_name: de
data_files:
- split: top
path: de/top.parquet
- split: high
path: de/high.parquet
- split: medium
path: de/medium.parquet
- split: low
path: de/low.parquet
- config_name: en
data_files:
- split: top
path: en/top.parquet
- split: high
path: en/high.parquet
- split: medium
path: en/medium.parquet
- split: low
path: en/low.parquet
- config_name: es
data_files:
- split: top
path: es/top.parquet
- split: high
path: es/high.parquet
- split: medium
path: es/medium.parquet
- split: low
path: es/low.parquet
- config_name: fr
data_files:
- split: top
path: fr/top.parquet
- split: high
path: fr/high.parquet
- split: medium
path: fr/medium.parquet
- split: low
path: fr/low.parquet
- config_name: id
data_files:
- split: top
path: id/top.parquet
- split: high
path: id/high.parquet
- split: medium
path: id/medium.parquet
- split: low
path: id/low.parquet
- config_name: it
data_files:
- split: top
path: it/top.parquet
- split: high
path: it/high.parquet
- split: medium
path: it/medium.parquet
- split: low
path: it/low.parquet
- config_name: ja
data_files:
- split: top
path: ja/top.parquet
- split: high
path: ja/high.parquet
- split: medium
path: ja/medium.parquet
- split: low
path: ja/low.parquet
- config_name: ko
data_files:
- split: top
path: ko/top.parquet
- split: high
path: ko/high.parquet
- split: medium
path: ko/medium.parquet
- split: low
path: ko/low.parquet
- config_name: ms
data_files:
- split: top
path: ms/top.parquet
- split: high
path: ms/high.parquet
- split: medium
path: ms/medium.parquet
- split: low
path: ms/low.parquet
- config_name: pt
data_files:
- split: top
path: pt/top.parquet
- split: high
path: pt/high.parquet
- split: medium
path: pt/medium.parquet
- split: low
path: pt/low.parquet
- config_name: ru
data_files:
- split: top
path: ru/top.parquet
- split: high
path: ru/high.parquet
- split: medium
path: ru/medium.parquet
- split: low
path: ru/low.parquet
- config_name: sw
data_files:
- split: top
path: sw/top.parquet
- split: high
path: sw/high.parquet
- split: medium
path: sw/medium.parquet
- split: low
path: sw/low.parquet
- config_name: te
data_files:
- split: top
path: te/top.parquet
- split: high
path: te/high.parquet
- split: medium
path: te/medium.parquet
- split: low
path: te/low.parquet
- config_name: th
data_files:
- split: top
path: th/top.parquet
- split: high
path: th/high.parquet
- split: medium
path: th/medium.parquet
- split: low
path: th/low.parquet
- config_name: vi
data_files:
- split: top
path: vi/top.parquet
- split: high
path: vi/high.parquet
- split: medium
path: vi/medium.parquet
- split: low
path: vi/low.parquet
- config_name: zh
data_files:
- split: top
path: zh/top.parquet
- split: high
path: zh/high.parquet
- split: medium
path: zh/medium.parquet
- split: low
path: zh/low.parquet
task_categories:
- text-generation
language:
- ar
- bn
- de
- en
- es
- fr
- id
- it
- ja
- ko
- ms
- pt
- ru
- sw
- te
- th
- vi
- zh
tags:
- mathematical-reasoning
- multilingual
- benchmark
- llm-evaluation
- math
PolyMath: Evaluating Mathematical Reasoning in Multilingual Contexts
Paper: https://arxiv.org/abs/2504.18428 Code: https://github.com/QwenLM/PolyMath Project Page: https://qwen-polymath.github.io/
PolyMath is a multilingual mathematical reasoning benchmark covering 18 languages and 4 easy-to-hard difficulty levels. Our benchmark ensures difficulty comprehensiveness, language diversity, and high-quality translation, making it a highly discriminative multilingual mathematical benchmark in the era of reasoning LLMs.
- 📈 Broad Difficulty Range: PolyMath defines and partitions mathematical difficulty across four levels using two core dimensions: Thought Depth and Knowledge Breadth, ranging from K-12 to Olympiad and advanced frontier mathematics, with 125 problems per language at each level.
- 🌍 Language Diversity: Each problem in PolyMath is available in 18 parallel language versions, encompassing over 75% of the world’s native speakers and major language families, ensuring diversity across both high-resource and low-resource languages.
- 🧑🏫 High-Quality Annotation: Each problem translation is calibrated by language experts, avoiding direct use of LLM-generated outputs and ensuring precise term and logical clarity.
📊 Main Results
The leaderboard is continuously updated! See https://qwen-polymath.github.io/#leaderboard
Sample Usage
You can easily load the dataset using the datasets library:
from datasets import load_dataset
# Load the English (en) split with 'high' difficulty
dataset = load_dataset("Qwen/PolyMath", "en", split="high")
print(dataset)
# Output: Dataset({
# features: ['problem_id', 'question', 'answer', 'level', 'difficulty', 'solution', 'image_paths'],
# num_rows: 125
# })
print(dataset[0])
# Output example: {'problem_id': 'en-high-0', 'question': '...', 'answer': '...', 'level': 'High', 'difficulty': 'Hard', 'solution': '...', 'image_paths': []}
# You can specify other languages and splits, e.g., for German (de) and 'top' difficulty:
# dataset_de_top = load_dataset("Qwen/PolyMath", "de", split="top")
For more detailed usage and evaluation scripts, please refer to the official GitHub repository.
📄 Citation
If you use PolyMath in your research, please cite us:
@article{wang2025polymath,
title={PolyMath: Evaluating Mathematical Reasoning in Multilingual Contexts},
author={Yiming Wang and Pei Zhang and Jialong Tang and Haoran Wei and Baosong Yang and Rui Wang and Chenshu Sun and Feitong Sun and Jiran Zhang and Junxuan Wu and Qiqian Cang and Yichang Zhang and Fei Huang and Junyang Lin and Fei Huang and Jingren Zhou},
journal={arXiv preprint arXiv:2504.18428},
year={2025},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2504.18428},
}