metadata

license: openrail
task_categories:
  - image-to-text

Dataset Description

There are two datasets here: raw_formulas and tokenized_formulas.

We crawled approximately 1 million uncleaned and untokenized latex formula image-text pairs from arxiv, resulting in the raw_formulas dataset. After cleaning and tokenizing the raw_formulas dataset, we obtained the tokenized_formulas dataset.

When rendering images corresponding to the formulas, the following external packages are required:

amsmath
amsfonts
amssymb
mathtools

Details about the raw_formulas dataset

We crawled latex formulas containing the following formula environments:

equation
align
align*
gather
gather*

The formulas do not contain the following content:

\label
%
\quad
\qquad
\vspace
\hspace
\resizebox
\scalebox
\rotatebox
\parbox
\fbox
\makebox
\raisebox
\addvspace
\hfill
\vfill
\textwidth
\textheight
\rule

Preprocessing details of the tokenized_formulas dataset

Cleaning

We removed some useless junk data from raw_formulas.
We removed overly complex formulas from raw_formulas:
Formulas with a render image aspect ratio greater than 0.8 were deleted.
Formulas with a character length greater than 200 were deleted.
The following content in raw_formulas formulas was removed:
- \tag,
- \text
- \begin{split}
- \end{split}
- \nonumber
- \notag
In raw_formulas, the equation, equation*, align, \[...\] formula environments were replaced with the align* formula environment, and the gather formula environment was replaced with the gather* environment.
We removed formulas containing custom macros from raw_formulas, only the following common custom macros were retained:
- \newcommand{\R}{\mathbb{R}}
- \newcommand{\N}{\mathbb{N}}
- \newcommand{\Z}{\mathbb{Z}}
- \newcommand{\Q}{\mathbb{Q}}
- \newcommand{\C}{\mathbb{C}}
- \newcommand{\avg}[1]{\left<#1\right>}
- \newcommand{\Deriv}[2]{\frac{\mathrm{d} #1}{\mathrm{d} #2}}
- \newcommand{\dd}{\mathrm{d}}
- \newcommand{\norm}[1]{\left\lVert#1\right\rVert}
- \newcommand{\abs}[1]{\left|#1\right|}
- \newcommand{\vect}[1]{\mathbf{#1}}

Tokenization

Substrings that match the following pattern are considered a word:

\begin{.*?}
\end{.*?}
\[A-Za-z]+

The following strings are also considered a word:

\[
\]
\{
\}
\_
\$
\&
\#
\%
\|
'
''
'''
''''
'^
''^
'''^
''''^