metadata
language:
- en
- hi
- zh
- pt
- sw
- he
license: apache-2.0
task_categories:
- multiple-choice
- question-answering
size_categories:
- n<1K
tags:
- culture
- multilingual
- evaluation
pretty_name: Multilingual CulturalBench
configs:
- config_name: default
data_files:
- split: train
path: multilingual_cultural_bench.csv
Multilingual CulturalBench (Translated Subset)
This dataset is a multilingual extension of the CulturalBench dataset ("Easy" subset). It contains 787 samples from the original benchmark, translated into five additional languages using Gemini-2.5-Flash.
Dataset Description
The original CulturalBench is designed to assess the cultural capabilities of Large Language Models (LLMs). This version extends the "Easy" subset (multiple-choice questions) by providing translations for the questions and options, enabling multilingual cultural evaluation.
Note: This is a partial subset (787 samples) of the original 1,230 "Easy" samples.
Languages
The dataset includes parallel data in the following languages:
- English (Original)
- Hindi
- Chinese (Simplified)
- Brazilian Portuguese (
braz_port) - Swahili
- Hebrew
Statistics
Total Samples: 787
Questions by Country
| Country | Count | Country | Count | Country | Count |
|---|---|---|---|---|---|
| China | 59 | Vietnam | 27 | United States | 20 |
| South Africa | 58 | United Kingdom | 25 | Peru | 19 |
| Japan | 53 | Brazil | 25 | Morocco | 17 |
| India | 46 | Bangladesh | 25 | Saudi Arabia | 17 |
| South Korea | 41 | Singapore | 23 | Australia | 15 |
| Iran | 37 | Lebanon | 22 | Pakistan | 14 |
| Hong Kong | 36 | Nigeria | 22 | France | 14 |
| Argentina | 35 | Chile | 22 | ||
| Germany | 32 | Taiwan | 22 | ||
| Nepal | 21 | Israel | 13 | ||
| Egypt | 20 | Canada | 7 | Total | 787 |
Dataset Structure
The dataset is a CSV file (multilingual_cultural_bench.csv) with 33 columns. Each row represents a single question with its translations and ground truth.
Fields
| Field | Description |
|---|---|
question_idx |
Unique identifier from the original CulturalBench. |
english_question |
The original question in English. |
english_option_[a-d] |
The four options in English. |
hindi_question |
Translated question in Hindi. |
hindi_option_[a-d] |
Translated options in Hindi. |
chinese_question |
Translated question in Chinese. |
chinese_option_[a-d] |
Translated options in Chinese. |
braz_port_question |
Translated question in Brazilian Portuguese. |
braz_port_option_[a-d] |
Translated options in Brazilian Portuguese. |
swahili_question |
Translated question in Swahili. |
swahili_option_[a-d] |
Translated options in Swahili. |
hebrew_question |
Translated question in Hebrew. |
hebrew_option_[a-d] |
Translated options in Hebrew. |
answer |
The correct option label (A, B, C, or D). |
country |
The country/culture the question pertains to. |
Sample Questions
Sample 1: South Korea (Dining Etiquette)
Question
| Language | Question |
|---|---|
| English | In Korean dining etiquette, what is a common practice regarding drinks and paying for the meal? |
| Hindi | कोरियाई भोजन शिष्टाचार में, पेय और भोजन के भुगतान के संबंध में एक सामान्य प्रथा क्या है? |
| Chinese | 在韩国的餐桌礼仪中，关于饮料和买单，常见的做法是什么？ |
| Brazilian Portuguese | Na etiqueta de refeições coreana, qual é uma prática comum em relação a bebidas e ao pagamento da refeição? |
Options
| Language | Option A | Option B | Option C | Option D |
|---|---|---|---|---|
| English | Everyone pays for their own meal and drinks. | Younger people pour drinks for the elders and the elders pay for the meal. | The person who is older always pays, regardless of who did the invitation | It's common to split the bill evenly amongst all diners, regardless of age. |
| Hindi | हर कोई अपने भोजन और पेय का भुगतान स्वयं करता है। | युवा लोग बड़ों के लिए पेय परोसते हैं और बड़े भोजन का भुगतान करते हैं। | जो व्यक्ति बड़ा होता है वह हमेशा भुगतान करता है, भले ही निमंत्रण किसने दिया हो। | उम्र की परवाह किए बिना, सभी मेहमानों के बीच बिल को समान रूप से विभाजित करना आम बात है। |
| Chinese | 每个人都支付自己的餐费和饮料费。 | 晚辈为长辈倒酒，长辈支付餐费。 | 无论谁邀请的，年长的人总是付钱。 | 无论年龄大小，所有用餐者平均分摊账单是很常见的。 |
| Brazilian Portuguese | Todos pagam sua própria refeição e bebidas. | Os mais jovens servem bebidas para os mais velhos e os mais velhos pagam a refeição. | A pessoa mais velha sempre paga, independentemente de quem fez o convite. | É comum dividir a conta igualmente entre todos os comensais, independentemente da idade. |
Correct Answer: Option B
Sample 2: Morocco (Greeting)
Question
| Language | Question |
|---|---|
| English | In Moroccan culture, how do people typically greet people they know on the street? |
| Swahili | Katika utamaduni wa Moroko, watu huwasalimu vipi watu wanaowajua barabarani? |
| Hebrew | בתרבות המרוקאית, כיצד אנשים מברכים בדרך כלל אנשים שהם מכירים ברחוב? |
| Chinese | 在摩洛哥文化中，人们通常如何问候街上的熟人？ |
Options
| Language | Option A | Option B | Option C | Option D |
|---|---|---|---|---|
| English | By saying "peace be upon you" | By saying "hello" | By hugging them | By giving a casual nod. |
| Swahili | Kwa kusema "amani iwe juu yako" | Kwa kusema "habari" | Kwa kuwakumbatia | Kwa kuinama kichwa kawaida. |
| Hebrew | באמירת "שלום עליכם" | באמירת "שלום" | בחיבוק שלהם | בקידה אגבית. |
| Chinese | 说“愿和平降临于你” | 说“你好” | 拥抱他们 | 随意地点头致意 |
Correct Answer: Option A
Usage
You can load this dataset using the Hugging Face datasets library:
from datasets import load_dataset
# Load the dataset
dataset = load_dataset("Lossfunk/Multilingual_CulturalBench")
# Access the 'train' split
train_data = dataset['train']
# Example: Accessing a Hindi question from the first sample
print(train_data[0]['hindi_question'])
print(train_data[0]['hindi_option_a'])