Belebele Multilingual Reading Comprehension Benchmark
Belebele is a parallel multiple-choice reading comprehension benchmark covering 122 language variants, with 900 questions per language built on FLORES-200 passages. Reproduce and annotate it in Potato.
About this dataset
Belebele is a machine reading comprehension benchmark released by Bandarkar et al. at Meta AI and published at ACL 2024. Every question is fully parallel across all languages, so the same passage and question set is translated into each variant for direct cross-lingual comparison.
The dataset contains 122 language variants spanning 115 distinct languages, 27 language families, and 29 scripts. Each variant has 900 questions for 109,800 questions in total. The questions draw on 488 distinct passages from FLORES-200, with one or two questions per passage. Passages average 79.1 words and 4.1 sentences.
Each question presents four answer options, exactly one of which is correct, recorded as a one-indexed correct_answer_num. The task evaluates passage-grounded reading comprehension across high-, medium-, and low-resource languages, and the English subset alone is hard enough to challenge strong language models.
The Potato config below reproduces the Belebele multiple-choice task as a radio-button annotation interface. Annotators read a passage, answer a four-option question, and add metadata such as difficulty and reasoning type.
- Language variants
- 122
- Distinct languages
- 115
- Questions per language
- 900
- Total questions
- 109,800
- Distinct passages
- 488 (from FLORES-200)
- Answer options
- 4 (1 correct)
Configuration Fileconfig.yaml
This Potato config reproduces the annotation task. Save it as config.yaml and run potato start config.yaml to try it.
# Multilingual Reading Comprehension (Belebele)
# Based on Bandarkar et al., ACL 2024
# Paper: https://arxiv.org/abs/2308.16884
# Dataset: https://github.com/facebookresearch/belebele
#
# This task presents a passage and a multiple-choice question with four
# answer options. The Belebele benchmark provides parallel questions across
# 122 languages, making it a key resource for evaluating cross-lingual
# reading comprehension. This configuration uses English passages.
#
# Answer Options:
# - A, B, C, D: Four possible answers; exactly one is correct
#
# Difficulty Levels:
# - EASY: Answer is explicitly stated in the passage
# - MEDIUM: Requires simple inference or paraphrasing
# - HARD: Requires complex reasoning or synthesis of multiple parts
#
# Reasoning Types:
# - FACTUAL: Answer is directly stated in the passage
# - INFERENTIAL: Answer requires drawing conclusions from the text
# - LEXICAL: Answer requires understanding vocabulary or paraphrasing
# - STRUCTURAL: Answer requires understanding passage organization
#
# Annotation Guidelines:
# 1. Read the passage thoroughly before looking at the question
# 2. Read the question and all four answer options carefully
# 3. Select the single best answer based on the passage content
# 4. Do not rely on external knowledge -- answers come from the passage
# 5. Rate the difficulty based on how much reasoning is required
# 6. Identify what type of reasoning is needed to arrive at the answer
#
# Important Notes:
# - Only one answer is correct per question
# - Some distractors may be partially correct or plausible
# - The passage contains all information needed to answer
annotation_task_name: "Multilingual Reading Comprehension (Belebele)"
task_dir: "."
data_files:
- sample-data.json
item_properties:
id_key: "id"
text_key: "passage"
output_annotation_dir: "annotation_output/"
output_annotation_format: "json"
port: 8000
server_name: localhost
annotation_schemes:
# Step 1: Select the correct answer
- annotation_type: radio
name: correct_answer
description: "Based on the passage, which answer is correct?"
labels:
- "A"
- "B"
- "C"
- "D"
keyboard_shortcuts:
"A": "1"
"B": "2"
"C": "3"
"D": "4"
tooltips:
"A": "Select if option A is the correct answer"
"B": "Select if option B is the correct answer"
"C": "Select if option C is the correct answer"
"D": "Select if option D is the correct answer"
# Step 2: Rate difficulty
- annotation_type: radio
name: difficulty
description: "How difficult is this question to answer correctly?"
labels:
- "easy"
- "medium"
- "hard"
tooltips:
"easy": "The answer is explicitly stated in the passage with minimal reasoning"
"medium": "The answer requires some inference or paraphrasing"
"hard": "The answer requires complex reasoning or synthesis of multiple passage parts"
# Step 3: Reasoning type
- annotation_type: radio
name: reasoning_type
description: "What type of reasoning is required to answer this question?"
labels:
- "factual"
- "inferential"
- "lexical"
- "structural"
tooltips:
"factual": "The answer is directly stated in the passage (fact lookup)"
"inferential": "The answer requires drawing a conclusion from passage information"
"lexical": "The answer requires understanding vocabulary, synonyms, or paraphrases"
"structural": "The answer requires understanding how the passage is organized"
annotation_instructions: |
You will be shown a passage followed by a question and four answer options (A, B, C, D). Your task is to:
1. Read the passage carefully.
2. Select the one correct answer based on the passage.
3. Rate how difficult the question is.
4. Identify what type of reasoning is needed.
Remember: base your answer only on the passage content, not on external knowledge.
html_layout: |
<div style="padding: 15px; max-width: 800px; margin: auto;">
<div style="background: #f0f9ff; border: 1px solid #bae6fd; border-radius: 8px; padding: 16px; margin-bottom: 16px;">
<strong style="color: #0369a1;">Passage:</strong>
<p style="font-size: 16px; line-height: 1.7; margin: 8px 0 0 0;">{{passage}}</p>
</div>
<div style="background: #fefce8; border: 1px solid #fde68a; border-radius: 8px; padding: 16px; margin-bottom: 16px;">
<strong style="color: #a16207;">Question:</strong>
<p style="font-size: 16px; line-height: 1.6; margin: 8px 0 0 0;">{{question}}</p>
</div>
<div style="display: grid; grid-template-columns: 1fr 1fr; gap: 10px;">
<div style="background: #f8fafc; border: 1px solid #e2e8f0; border-radius: 8px; padding: 12px;">
<strong style="color: #475569;">A:</strong> {{option_a}}
</div>
<div style="background: #f8fafc; border: 1px solid #e2e8f0; border-radius: 8px; padding: 12px;">
<strong style="color: #475569;">B:</strong> {{option_b}}
</div>
<div style="background: #f8fafc; border: 1px solid #e2e8f0; border-radius: 8px; padding: 12px;">
<strong style="color: #475569;">C:</strong> {{option_c}}
</div>
<div style="background: #f8fafc; border: 1px solid #e2e8f0; border-radius: 8px; padding: 12px;">
<strong style="color: #475569;">D:</strong> {{option_d}}
</div>
</div>
</div>
allow_all_users: true
instances_per_annotator: 50
annotation_per_instance: 2
allow_skip: true
skip_reason_required: false
Sample Datasample-data.json
[
{
"id": "belebele_001",
"passage": "Photosynthesis is the process by which green plants and certain other organisms transform light energy into chemical energy. During photosynthesis, plants capture light energy from the sun and use it to convert water and carbon dioxide into oxygen and glucose. The glucose is used as food by the plant, while the oxygen is released into the atmosphere. This process primarily takes place in the leaves of the plant, specifically in organelles called chloroplasts that contain the green pigment chlorophyll.",
"question": "Where does photosynthesis primarily occur within a plant?",
"option_a": "In the roots where water is absorbed",
"option_b": "In the stem where nutrients are transported",
"option_c": "In the leaves, within chloroplasts",
"option_d": "In the flowers during pollination",
"language": "English"
},
{
"id": "belebele_002",
"passage": "The Renaissance was a cultural movement that began in Italy in the late 14th century and later spread throughout Europe. It marked a period of renewed interest in classical Greek and Roman art, literature, and philosophy. Key figures of the Renaissance include Leonardo da Vinci, Michelangelo, and Raphael in the visual arts, and Petrarch and Boccaccio in literature. The invention of the printing press by Johannes Gutenberg around 1440 played a crucial role in disseminating Renaissance ideas across the continent.",
"question": "What technology helped spread Renaissance ideas across Europe?",
"option_a": "The compass",
"option_b": "The printing press",
"option_c": "The telescope",
"option_d": "The steam engine",
"language": "English"
}
]
// ... and 8 more itemsGet This Design
Clone or download from the repository
Quick start:
git clone https://github.com/davidjurgens/potato-showcase.git cd potato-showcase/text/cross-lingual/belebele-reading-comprehension potato start config.yaml
Dataset & paper
Bandarkar et al., ACL 2024
Citation (BibTeX)
@inproceedings{bandarkar-etal-2024-belebele,
title = "The Belebele Benchmark: a Parallel Reading Comprehension Dataset in 122 Language Variants",
author = "Bandarkar, Lucas and Liang, Davis and Muller, Benjamin and Artetxe, Mikel and Shukla, Satya Narayan and Husa, Donald and Goyal, Naman and Krishnan, Abhinandan and Zettlemoyer, Luke and Khabsa, Madian",
booktitle = "Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
month = aug,
year = "2024",
address = "Bangkok, Thailand",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.acl-long.44/",
pages = "749--775"
}Details
Annotation Types
Domain
Use Cases
Tags
Found an issue or want to improve this design?
Open an IssueRelated Designs
Machine Comprehension Using Commonsense Knowledge
Multiple-choice reading comprehension requiring commonsense reasoning over narrative texts, selecting the best answer and providing reasoning. Based on SemEval-2018 Task 11.
#HashtagWars - Learning a Sense of Humor
Humor ranking of tweets submitted to Comedy Central's @midnight #HashtagWars, classifying comedic quality. Based on SemEval-2017 Task 6.
ArgSciChat Scientific Argumentation Dialogue
Annotation of argumentative dialogues about scientific papers based on the ArgSciChat dataset. Annotators label dialogue turns for argument components (claim, evidence, rebuttal) and assess argument quality dimensions such as clarity, relevance, and persuasiveness.