| --- |
| annotations_creators: |
| - expert-generated |
| language: |
| - en |
| - am |
| - ar |
| - zh |
| - ka |
| - el |
| - hi |
| - id |
| - it |
| - ja |
| - ky |
| - es |
| - pt |
| - yo |
| - tl |
| - zu |
| - th |
| - tr |
| task_categories: |
| - question-answering |
| - text-classification |
| pretty_name: Macaron |
| arxiv: 2602.10732 |
| tags: |
| - benchmark |
| - evaluation |
| - multilingual |
| - multicultural |
| - reasoning |
| - template-based |
| configs: |
| - config_name: MCQ |
| data_files: macaron_mcq.csv |
| - config_name: True-False |
| data_files: macaron_truefalse.csv |
| --- |
| |
| # Macaron |
|
|
| [[Paper](https://huggingface.co/papers/2602.10732)] |
|
|
| **Macaron** is a controlled, human-written benchmark for **multilingual and multicultural reasoning** created with a **template-first** approach. |
| Each example is **scenario-aligned** across **English** and a **local language**, enabling controlled comparison of reasoning under culturally grounded premises. |
|
|
| --- |
|
|
| ## At a glance |
|
|
| | Configuration | Rows | Description | |
| |--------------|------|-------------| |
| | MCQ | 1,977 | Bilingual multiple-choice questions (English + local language) | |
| | True-False | 3,954 | Bilingual verification statements derived from MCQs | |
| | Templates | 100 | Reusable templates with reasoning and cultural metadata | |
|
|
| > Counting each language-specific instance separately (English + local), the benchmark contains **11,862 evaluation instances**. |
|
|
| --- |
|
|
| ## Supported tasks |
|
|
| - Multiple-choice question answering |
| - Binary classification / verification (True/False) |
|
|
| --- |
|
|
| ## Coverage |
|
|
| Macaron provides controlled coverage across **languages, cultural contexts, reasoning types, and cultural aspects**. |
| All instances are **scenario-aligned** across English and a local language. |
|
|
| ### Languages and cultural contexts |
|
|
| The benchmark spans **20 cultural contexts**, each paired with **English** and one primary local language. |
|
|
| | Country / Context | Local language | |
| |------------------|----------------| |
| | Brazil | Brazilian Portuguese | |
| | China | Chinese | |
| | Egypt | Egyptian Arabic | |
| | Ethiopia | Amharic | |
| | Georgia | Georgian | |
| | Greece | Greek | |
| | India | Hindi | |
| | Indonesia | Indonesian | |
| | Italy | Italian | |
| | Japan | Japanese | |
| | Kyrgyzstan | Kyrgyz | |
| | Mexico | Mexican Spanish | |
| | Morocco | Moroccan Arabic | |
| | Nigeria | Yoruba | |
| | Philippines | Tagalog | |
| | South Africa | Zulu | |
| | Thailand | Thai | |
| | Tunisia | Tunisian Arabic | |
| | Turkey | Turkish | |
| | Yemen | Yemeni Arabic | |
|
|
| --- |
|
|
| ### Dataset size by context |
|
|
| Each multiple-choice question (MCQ) produces: |
| - **1 MCQ row** |
| - **2 True-False rows** |
|
|
| Each row contains **both English and local-language text**. |
|
|
| | Country / Context | MCQ rows | True-False rows | Evaluation instances (EN + Local) | |
| |------------------|----------|-----------------|----------------------------------| |
| | Brazil | 100 | 200 | 600 | |
| | China | 97 | 194 | 582 | |
| | Egypt | 99 | 198 | 594 | |
| | Ethiopia | 98 | 196 | 588 | |
| | Georgia | 99 | 198 | 594 | |
| | Greece | 100 | 200 | 600 | |
| | India | 100 | 200 | 600 | |
| | Indonesia | 95 | 190 | 570 | |
| | Italy | 98 | 196 | 588 | |
| | Japan | 99 | 198 | 594 | |
| | Kyrgyzstan | 100 | 200 | 600 | |
| | Mexico | 99 | 198 | 594 | |
| | Morocco | 100 | 200 | 600 | |
| | Nigeria | 95 | 190 | 570 | |
| | Philippines | 99 | 198 | 594 | |
| | South Africa | 100 | 200 | 600 | |
| | Thailand | 99 | 198 | 594 | |
| | Tunisia | 100 | 200 | 600 | |
| | Turkey | 100 | 200 | 600 | |
| | Yemen | 100 | 200 | 600 | |
|
|
| --- |
|
|
| ### Reasoning types |
|
|
| Each template and derived instance is tagged with one or more **reasoning types**. |
|
|
| | Reasoning type | Description | |
| |---------------|-------------| |
| | Mathematical Reasoning | Numerical computation and comparison | |
| | Commonsense Reasoning | Everyday plausibility and typical situations | |
| | Causal Reasoning | Cause–effect relations | |
| | Temporal Reasoning | Time, order, calendars | |
| | Logical Reasoning | Deduction, implication, and analogy | |
| | Spatial Reasoning | Geographic and spatial relations | |
| | Multi-hop Reasoning | Composition of two or more inference steps | |
|
|
| --- |
|
|
| ### Cultural aspects |
|
|
| Templates are tagged with one or more **cultural aspects**, covering **22 domains of everyday life**: |
|
|
| - agriculture |
| - brands and commerce |
| - cities and landmarks |
| - death and funerals |
| - education |
| - events and festivals |
| - famous people |
| - fashion and media |
| - folklore and folktales |
| - food and cuisine |
| - language and communication |
| - literature and written works |
| - music and art |
| - naming |
| - objects and units |
| - politics and governance |
| - relationships |
| - social customs |
| - sports |
| - time |
| - transportation |
| - socio-religious aspects of life |
|
|
| Both `reasoning_category` and `cultural_aspect` fields are **multi-label**, stored as comma-separated strings in the CSV files. |
|
|
| --- |
|
|
| ## Intended use |
|
|
| Macaron is intended for: |
|
|
| - Zero-shot and few-shot evaluation of multilingual large language models |
| - Cross-lingual robustness analysis using scenario-aligned English and local-language inputs |
| - Diagnostic analysis by reasoning type and cultural domain |
|
|
| Not recommended uses: |
|
|
| - Training and testing on the same benchmark |
| - Drawing broad conclusions about entire cultures, countries, or languages |
|
|
| --- |
|
|
| ## How to load |
|
|
| ```python |
| from datasets import load_dataset |
| |
| ds_mcq = load_dataset("AlaaAhmed2444/Macaron", "MCQ") |
| ds_tf = load_dataset("AlaaAhmed2444/Macaron", "True-False") |
| ds_tpl = load_dataset("AlaaAhmed2444/Macaron", "Templates") |
| ``` |
| --- |
|
|
| ## Ethical considerations and limitations |
|
|
| - Cultural coverage is necessarily coarse: each cultural context is represented by one primary local language and does not capture within-country diversity or dialect continua. |
| - The benchmark focuses on controlled reasoning formats (multiple-choice and True/False), which do not reflect open-ended dialogue or interactive reasoning settings. |
| - Results should not be interpreted as representing full cultural or linguistic diversity, but rather as performance on a controlled, template-based evaluation. |
|
|
| --- |
|
|
| ## Citation |
|
|
| If you use Macaron, please cite the accompanying paper: |
|
|
| ```bibtex |
| @misc{elsetohy2026macaron, |
| title={Macaron: Controlled, Human-Written Benchmark for Multilingual and Multicultural Reasoning via Template-Filling}, |
| author={Elsetohy, Alaa and Hadhoud, Sama and Wibowo, Haryo Akbarianto and Whitehouse, Chenxi and Winata, Genta Indra and Koto, Fajri and Aji, Alham Fikri}, |
| year={2026}, |
| eprint={2602.10732}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.CL}, |
| url={https://arxiv.org/abs/2602.10732} |
| } |
| ``` |
