ScienceQA Multimodal Reasoning
Multimodal science question answering with chain-of-thought reasoning, based on ScienceQA (Lu et al., NeurIPS 2022). Annotators answer multiple-choice science questions that may include images, provide chain-of-thought explanations, and categorize the science domain.
About this dataset
ScienceQA is a multimodal science question answering benchmark from Lu and colleagues at NeurIPS 2022. Its distinguishing feature is that most questions come with a grounding lecture and a written explanation, supporting chain-of-thought reasoning rather than answer-only prediction.
The questions are multiple-choice items drawn from elementary and high-school science curricula. Many pair a question with an image such as a diagram, map, or photo, so answering can require reading both text and a visual.
The dataset has 21,208 questions spanning 3 subjects, 26 topics, and 127 categories; 48.7% include an image, 83.9% include a lecture, and 90.5% include an explanation. Annotating the reasoning, not just the answer, is what makes it useful for studying explanations.
The Potato config below reproduces the task with a radio for the answer choice, a text field for the explanation, and a multiselect for the reasoning skills involved. Use it to extend ScienceQA or to collect explained answers for your own questions.
- Released
- NeurIPS 2022
- Questions
- 21,208 (multiple choice)
- With image context
- 10,332 (48.7%)
- Coverage
- 3 subjects, 26 topics, 127 categories
- With lectures
- 83.9%
- With explanations
- 90.5% (chain-of-thought)
Configuration Fileconfig.yaml
This Potato config reproduces the annotation task. Save it as config.yaml and run potato start config.yaml to try it.
# ScienceQA Multimodal Reasoning
# Based on Lu et al., NeurIPS 2022
# Paper: https://arxiv.org/abs/2209.09513
# Dataset: https://scienceqa.github.io/
#
# Multimodal science question answering with chain-of-thought reasoning.
# Questions span natural science, social science, and language science,
# and may include images (diagrams, charts, photos). Annotators select
# the correct answer, provide a chain-of-thought explanation, and
# categorize the science domain.
#
# Answer Options:
# - A, B, C, D: Four possible answers; exactly one is correct
#
# Science Domains:
# - Natural Science: Biology, physics, chemistry, earth science
# - Social Science: History, geography, economics, civics
# - Language Science: Grammar, vocabulary, reading comprehension
#
# Annotation Guidelines:
# 1. Examine any image provided carefully
# 2. Read the question and all four options
# 3. Select the correct answer
# 4. Write a step-by-step chain-of-thought explanation
# 5. Categorize the question into its science domain(s)
annotation_task_name: "ScienceQA Multimodal Reasoning"
task_dir: "."
data_files:
- sample-data.json
item_properties:
id_key: "id"
text_key: "text"
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: answer
description: "Select the correct answer to the science question."
labels:
- "A"
- "B"
- "C"
- "D"
keyboard_shortcuts:
"A": "1"
"B": "2"
"C": "3"
"D": "4"
tooltips:
"A": "Select option A"
"B": "Select option B"
"C": "Select option C"
"D": "Select option D"
# Step 2: Chain-of-thought explanation
- annotation_type: text
name: chain_of_thought
description: "Provide a step-by-step chain-of-thought explanation for your answer."
textarea: true
required: false
placeholder: "Step 1: ... Step 2: ... Therefore, the answer is ..."
# Step 3: Science domain categorization
- annotation_type: multiselect
name: science_domain
description: "Which science domain(s) does this question belong to? Select all that apply."
labels:
- "Natural Science"
- "Social Science"
- "Language Science"
tooltips:
"Natural Science": "Biology, physics, chemistry, earth science, astronomy"
"Social Science": "History, geography, economics, civics, sociology"
"Language Science": "Grammar, vocabulary, reading comprehension, linguistics"
annotation_instructions: |
You will answer science questions from the ScienceQA benchmark. Questions may include images.
For each item:
1. If an image is provided, examine it carefully.
2. Read the question and all four answer options (A-D).
3. Select the single correct answer.
4. Write a chain-of-thought explanation showing your reasoning step by step.
5. Categorize the question into its science domain(s).
Chain-of-Thought Tips:
- Break your reasoning into clear, numbered steps.
- Reference specific information from the image or question.
- Explain why incorrect options are wrong when helpful.
- Conclude with "Therefore, the answer is [letter]."
html_layout: |
<div style="padding: 15px; max-width: 800px; margin: auto;">
<div style="background: #e8eaf6; padding: 8px 15px; border-radius: 8px; margin-bottom: 16px;">
<strong>Subject:</strong> {{subject}}
</div>
<div style="text-align: center; margin-bottom: 16px;">
<img src="{{image_url}}" style="max-width: 100%; max-height: 400px; border: 1px solid #ddd; border-radius: 8px;" />
</div>
<div style="background: #f0f9ff; border: 1px solid #bae6fd; border-radius: 8px; padding: 16px; margin-bottom: 16px;">
<strong style="color: #0369a1;">Question:</strong>
<p style="font-size: 16px; line-height: 1.7; margin: 8px 0 0 0;">{{text}}</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": "sciqa_001",
"text": "Which of these organisms is a producer in the food chain shown in the diagram?",
"image_url": "https://example.com/scienceqa/image_001.jpg",
"option_a": "Grass",
"option_b": "Rabbit",
"option_c": "Fox",
"option_d": "Eagle",
"subject": "Biology"
},
{
"id": "sciqa_002",
"text": "Based on the weather map, which city is most likely to experience rain tomorrow?",
"image_url": "https://example.com/scienceqa/image_002.jpg",
"option_a": "City A, which is under a high-pressure system",
"option_b": "City B, which is near a cold front",
"option_c": "City C, which is in a clear area",
"option_d": "City D, which is far from any weather systems",
"subject": "Earth Science"
}
]
// ... 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/multimodal/scienceqa-multimodal-reasoning potato start config.yaml
Dataset & paper
Lu et al., NeurIPS 2022
Citation (BibTeX)
@inproceedings{lu2022learn,
title = "Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering",
author = "Lu, Pan and Mishra, Swaroop and Xia, Tony and Qiu, Liang and Chang, Kai-Wei and Zhu, Song-Chun and Tafjord, Oyvind and Clark, Peter and Kalyan, Ashwin",
booktitle = "Advances in Neural Information Processing Systems",
year = "2022",
url = "https://arxiv.org/abs/2209.09513"
}Details
Annotation Types
Domain
Use Cases
Tags
Found an issue or want to improve this design?
Open an IssueRelated Designs
MMBench: Multimodal Vision-Language Benchmark
MMBench is a multiple-choice benchmark of about 2,974 questions testing vision-language models across 20 ability dimensions (Liu et al., ECCV 2024). This Potato config reproduces image-question answering with skill tagging.
SayCan: Grounding Language in Robotic Affordances
SayCan grounds a large language model in learned robot skills so a robot can carry out long-horizon natural-language instructions. This Potato config reproduces the human evaluation of its generated action plans.
DocBank Document Layout Detection
Document layout analysis benchmark (Li et al., COLING 2020). Detect and classify document elements including titles, abstracts, paragraphs, figures, tables, and captions.