|
|
--- |
|
|
license: apache-2.0 |
|
|
task_categories: |
|
|
- question-answering |
|
|
language: |
|
|
- en |
|
|
tags: |
|
|
- math |
|
|
--- |
|
|
# Dataset Card for MSE-text-img-dataset |
|
|
|
|
|
<!-- Provide a quick summary of the dataset. --> |
|
|
|
|
|
We have created a custom dataset |
|
|
that is extracted as a subset of the Math Stack Exchange (MSE) |
|
|
dataset. This text-image dataset contains 64,860 questions with |
|
|
their respective list of answers, scores, acceptance marking, |
|
|
and image versions of each question and answer generated |
|
|
from the stored text with embedded LaTeX math markup. |
|
|
In this dataset there are 117,380 answers in total, with 1.81 |
|
|
answers per question on average. Each image was rendered |
|
|
from the question and answer text as it would appear in an |
|
|
equivalent PDF document, and the question answers selected |
|
|
were the first 64,860 questions and equivalent answers found |
|
|
in the original MSE dataset. Each question and associated |
|
|
answers are stored in JSON, with a JSON file for the test, train, |
|
|
and validation datasets. The text-image dataset has been split |
|
|
by 90% for training, 5% for testing, and 5% for validation. |
|
|
There are three image directories for the test, training, and |
|
|
validation images respectively, with each question’s image |
|
|
directory labeled by question ID from the JSON files, and |
|
|
each answer image by the answer ID with the question image |
|
|
simply labeled as ”question”. |
