Image-to-Markup Generation with Coarse-to-Fine Attention
Paper • 1609.04938 • Published • 2
Datasets:
Tasks:
Image-to-Text
Formats:
parquet
Size:
100K - 1M
ArXiv:
Tags:
im2latex
im2latex-100k
image-to-markup
mathematical-expression-recognition
latex
optical-character-recognition
License:
formula stringlengths 1 948 | filename stringlengths 14 14 | image imagewidth (px) 320 320 | preprocessing stringclasses 3
values |
|---|---|---|---|
L_{E-H}=\epsilon _{abcd}R^{ab}e^{c}e^{d}+dB, \label{EH} | 3c8a88ae01.png | none | |
- { \textstyle \frac { 1 } { 2 } } \omega ^ { + a b } _ { \hat { 0 } } \sigma _ { a b } \epsilon _ { I } = \gamma ^ { i } ( - { \textstyle \frac { i } { 4 } } e ^ { 3 \phi } \partial _ { \hat { \imath } } \overline { \lambda } \gamma ^ { 0 } \epsilon _ { I } ) \, . | 15c867a6fd.png | tokenized | |
- \kappa ^ { 2 } g ^ { 2 } \alpha ^ { \prime } \left( \frac { 7 1 1 6 8 } { 5 9 0 4 9 } \right) ( A ^ { \mu } A _ { \mu } ) ^ { 2 } . | 730c5a7d8e.png | normalized | |
\Phi _ A ^ { a b c } = 2 \varepsilon ^ { a \underline { b } } \rho _ A ^ { \underline { c } } | 77698e6a46.png | tokenized | |
\int _ { C } T _ { \mu \nu } k ^ { \mu } k ^ { \nu } d \lambda \ge 0 | 689c90715b.png | normalized | |
{\cal W}_{\rm PM}(\phi ) =\frac{m^2}{4\lambda}\,\phi-\frac{\lambda}{3}\,\phi^3\,,\label{PS} | 3261c47dbf.png | none | |
{ \frac { 2 m } { N } } - 2 m _ { 1 } = r - \ell \; . | f999974fcf.png | normalized | |
p(\beta,L)=\rho(\beta,L)= \frac{1}{L^{2}}f'\left(\frac{\beta}{L}\right)\;. | 490579c383.png | none | |
\tilde { S } _ { 0 } ^ { L } \left [ A _ { \mu } ^ { a } , B _ { a } ^ { \mu \nu } \right ] = \frac { 1 } { 2 } | 622144b02c.png | tokenized | |
d s ^ { 2 } = h ( l ) d t ^ { 2 } - \frac { d l ^ { 2 } } { h ( l ) } - d y ^ { 2 } - d z ^ { 2 } | 4f28beb022.png | normalized | |
(\stackrel{(0)}{S},\stackrel{(0)}{S})=0\longrightarrow(S,S)=0 \label{fullmasterequation}. | 8dde546d93.png | none | |
c=\frac{1}{1-\hat{u}}~~, \hspace{1cm}b=64 e^{2 \pi i\tau}.\label{15} | 29bd63f199.png | none | |
A _ { l } = \sum _ { k } \alpha _ { l k } ^ { * } a _ { k } - \beta _ { l k } ^ { * } a _ { k } ^ { \dagger } \; . | 62eba0fb07.png | normalized | |
\varphi(n)=\frac{\sin\vartheta n}{\sin\vartheta} \ \ . | 64496dd5d8.png | none | |
U^{\dagger}(\Lambda)\vec{\jmath}\,(P)U(\Lambda)={\cal R}_W(\Lambda,P)\vec{\jmath}. | 755a918771.png | none | |
\delta \left( Z _ { \; \; b _ { 1 } } ^ { a _ { 0 } } { \cal P } _ { a _ { 0 } } \right) = - D _ { \; \; b _ { 1 } } ^ { a _ { 1 } } \pi _ { a _ { 1 } } . | 65563600ef.png | normalized | |
0 = \int _ { K 3 } \eta _ { I J } F ^ { I } \wedge F ^ { J } = m ^ { A I } m ^ { B J } \rho _ { A B } \, \eta _ { I J } \ , | 154c1f75c2.png | normalized | |
n ( x ) = \Psi ^ { \dagger } ( x ) \Psi ( x ) , | 20cbba4e5a.png | tokenized | |
S _ { \mathrm { t o t a l } } = S + S _ { \mathrm { s o u r c e } } ~ , | 6495600e1f.png | normalized | |
\hat { \hat { \epsilon } } ^ { \hat { \hat { \mu } } _ { 0 } \ldots \hat { \hat { \mu } } _ { 1 0 } } = 1 . | 5defd0d234.png | normalized | |
{ \cal N } _ { A B } = \frac { \partial } { \partial \bar X ^ A } \frac { \partial } { \partial \bar X ^ B } \bar F \ . | 4382c7358f.png | tokenized | |
n_{i}\longrightarrow n_{i}+q_{i}m,\;\;m\in{\bf Z}\label{3.210} | 495140179c.png | none | |
m _ t / v \sim m _ b / w \sim ( 1 8 0 G e V / 2 4 6 G e V ) . | 52b4e85f58.png | tokenized | |
\bar { K } ( z ; i _ { s } ) \bar { Z } ( z ) G ( i _ { s } ) = Z ( z ^ { \prime } ) ^ { t } \, , | 18a0b011ab.png | tokenized | |
( { \bf u } _ { I } , [ D _ { a } , \Delta _ { f } ] { \bf u } _ { J } ) = 4 C _ { ( I J ) a } | 705378be90.png | normalized | |
\{ Q ^ a _ \alpha , Q ^ b _ \beta \} = 2 \omega ^ { a b } \gamma _ { \alpha \beta } ^ \mu P _ \mu + \gamma _ { \alpha \beta } ^ \mu Z ^ { a b } _ \mu | 7d13cc7b51.png | tokenized | |
m_l/M={\rm Log} \left( \frac {F_l}{F_0}\right) .\label{mass} | 25a3d6b46b.png | none | |
T^{(1,\dots ,N)}_{k}\left( \theta _{1},\dots ,\theta _{N}\right) | 2a1d9ee0c1.png | none | |
{\Pi}^{(-N)}=p^{(-N)} [ {\bf 1} - {\pi}^{(j^{(-N)})}], | 12d7c2d6c5.png | none | |
[ D, Q_{+{1\over 2}} ] = {1\over 2} Q_{+{1\over 2}} \ , \qquad [ D,S_{-{1\over 2}} ] = -{1\over 2} S_{-{1\over 2}} \ . | 4b452efb15.png | none | |
\label{n5.9}c_0=\int_{\bar{B}}{\bar h}^{1/2} d^2x=\mbox{Vol}~\bar{\cal B}, | 5d0d6d68cc.png | none | |
w = \underbrace { r _ { \epsilon } \cdots r _ { 0 } r _ { 1 } } _ { l } , | 772bdb8815.png | tokenized | |
r(n;s)=\int_s^\infty (x-s) \psi(q_0,\cdots,q_n)\, dx\label{r_integral} | 5161b969da.png | none | |
\eta _{2}\,=\,\frac{1\,-\,{\frac{\partial _{+}\,\hat{\varphi}}{\partial_{-}\,\hat{\varphi}}}}{1\,+\,{\frac{\partial _{+}\,\hat{\varphi}}{\partial_{-}\,\hat{\varphi}}}}\;\;. | 21461a3038.png | none | |
\label{symm:metr1} ds^2 = dy^A dy^B \eta_{AB} | 68d69adf99.png | none | |
M _ { \psi } ^ { 2 } = { \frac { d ^ { 2 } } { ( \pi \alpha ^ { \prime } ) ^ { 2 } } } \ , | 5b097371e7.png | normalized | |
\hat X_{i} \equiv\sqrt{ X_{i}^{2}+r_{0}^{2}}, \ \ \ X_{i}=(X,Y,P). | 18723c2065.png | none | |
\left( b + a \nabla ^ { 2 } \right) \nabla ^ { 2 } G ( \bar { x } , \bar { y } ) \, = \, \delta ^ { 2 } ( | 3a1b30f7bd.png | normalized | |
N t _ { 1 1 } \equiv k \ { \rm m o d } \ N \, . | 5b5eb83840.png | tokenized | |
\label{breflect}
\eta\sim \lambda_1\, e^{-i\omega t}\left(e^{ikx}+K_B\, e^{-ikx}\right) , | d698e7cd31.png | none | |
\label{cyl3} \lambda_i(\tau,x+2\pi) = \lambda_{Q(i)}(\tau,x)~. | 340f3b32e3.png | none | |
X _ \tau = F _ 1 X _ \sigma - F _ 2 X _ \sigma - F _ 3 X _ \tau , | 7236c5e44b.png | tokenized | |
\label{density_HO}\rho(q)=1,\;\; \rho(Q,t)={1\over 2\pi \sin t}. | 10a2d8959a.png | none | |
F _ { i } = F _ { i } ( \phi ) \ ; \ G _ { i } = G _ { i } ( \phi ) \ ; \ D _ { i } = D _ { i } ( \phi ) | 13e0a96d36.png | normalized | |
( d s ) ^ { 2 } = - [ d \tau ^ { 2 } - ( M \tau ) ^ { 2 } d x ^ { 2 } ] ; \hspace { 0 . 3 c m } 0 \le \tau < \infty | 5707afe575.png | normalized | |
\sum _ { i } { \cal C } _ { i } N _ { i } ^ { 2 } = \sum _ { i \neq j } { \cal L } _ { i j } | N _ { i } N _ { j } | + \sum _ { i } { \cal K } _ { i } N _ { i } ^ { 2 } . | 4c5f67d242.png | normalized | |
{\tilde D}_{m}= D_{m}-\frac{1}{2}m\Gamma_{m} | cd3df81d00.png | none | |
D = - ( \partial _ \mu + i A _ \mu ) ( \partial _ \mu + i A _ \mu ) + \sigma . | 4131a87e8b.png | tokenized | |
\Delta \sigma ( A ) = \Delta \sigma \left( 1 + A ^ { 2 } / 4 m ^ { 4 } ( \Delta \sigma ) ^ { 4 } \right) ^ { 1 / 2 } \ . | 1d84bfb0a6.png | normalized | |
{ Q } ( \iota _ * { F } ) = { Q } ( \iota _ * { E } ) + { Q } ( \iota _ * { G } ) \, . | 2d402da897.png | tokenized | |
G _ F = M _ { \rm W } ^ { - 2 } \sim 1 0 ^ { - 5 } \ \ \mbox { G e V } ^ { - 2 } | 7afe963f09.png | tokenized | |
U_{4\pi} = 4(\delta_{ij} \delta_{kl}+ cyclic) f^{\prime\prime}(v^2) | 468de1bc01.png | none | |
E ^ \prime = - B G ^ { - 1 } B - B = { B _ { 1 2 } ^ 2 \over \det ( G ) } ~ G - B ~ . | 41bb18acbf.png | tokenized | |
\phi _ i = x _ i / l _ s ^ 2 ~ , ~ \phi ^ 8 = x _ { 1 1 } / l _ s ^ 2 ~ , ~ u = v / l _ s ^ 2 , | 4b4fcc7892.png | tokenized | |
{ \cal P } _ { a } = p _ { a } \qquad { \cal J } _ { a } = \epsilon _ { a b c } x ^ { b } p ^ { c } + J _ { a } \, , | 4b14213a34.png | normalized | |
r _ { - } = - q _ { - } \circ \left( q _ { - } - \bar { M } \circ q _ { + } \right) ^ { - 1 } . | 4fda75d93a.png | normalized | |
\hat { c } _ { - \alpha } = e ^ { - i q \alpha } \sum _ { \beta \in \Lambda _ { R } } | \beta + \bar { p } > < \beta + \bar { p } | | 3a523a6698.png | tokenized | |
W _ { \perp } = e x p \{ \frac { i } { 2 } \operatorname { t a n } ^ { - 1 } \frac { p _ { \perp } } { m } \mathrm { \boldmath ~ \gamma ~ } _ { \perp } \cdot { \bf p } _ { \perp } \} | 27d084c7f2.png | normalized | |
[ { \cal E } ^ \alpha _ i ( x ) , A ^ \beta _ j ( x ' ) ] = i \delta ^ { \alpha \beta } \delta _ { i j } \delta ( x - x ' ) , | 5620a21fd2.png | tokenized | |
E _ c = - 2 \frac { n l r _ + ^ { n - 1 } V o l ( \Sigma _ n ) } { 1 6 \pi G R } . | 1a18ff0317.png | tokenized | |
\operatorname { l n } \, [ \alpha _ { a ( M _ { T } ) } / \alpha _ { a ( M _ { Z } ) } ] , \, \, \operatorname { l n } \, [ \alpha _ { a ( M _ { X } } ) / \alpha _ { a ( M _ { T } ) } ] | 6f23402f3b.png | normalized | |
| \Phi \rangle = | \Phi ^ { ( 0 ) } \rangle + a _ { _ U } | \Phi ^ { ( 1 ) } \rangle \, , | 4afee14dde.png | tokenized | |
u_a v^a = 0, \hspace{1.5em} v_a v^a = -2.%\label{U_VNorm} | 664e7c7be4.png | none | |
\lambda\equiv\tau_2/\tau_1=\theta g_{\rm YM}^2/8\pi^2~. | 1b1bea9d9a.png | none | |
\left| \left. b,\frac{1}{2},a,\frac{1}{2}\right| l,k-l,n\right\rangle \: . | 22659be215.png | none | |
x_d^{(1/2)}(u,\xi)={x(u,\xi)\,x(u+\omega_{1},\xi)\over x(\omega_{1},\xi)}. | 3b7c6d1cf7.png | none | |
G \left( \xi \right) = \frac { 3 } { 8 } \xi ^ { - 4 } - \frac { 7 } { 8 } \, \frac { 1 } { 1 2 0 } + O \left( \xi ^ { 2 } \right) , | 6c8f12167c.png | normalized | |
S _ { \mathrm { H } } ^ { 2 } - S _ { \mathrm { \Lambda } } ^ { 2 } = S _ { \mathrm { B H } } ( 2 S _ { \mathrm { B V } } - S _ { \mathrm { B H } } ) . | 3c05985550.png | normalized | |
a _ { q } = a f ( \hat { n } ) , ~ ~ ~ a _ { q } ^ { \dag } = f ( \hat { n } ) a ^ { \dag } | 4640fe556f.png | normalized | |
\Pi^{(2)}_{PV}(k^2,m)=\Pi^{(1)}(k^2,m)-\frac{g^2}{4\pi} sgn(M)\ .\label{2.9} | 7e92ebb935.png | none | |
[ e _ { a } , A _ { b } ^ { i } ( \vec { x } ) ] _ { D } = g c _ { a b c } A _ { c } ^ { i } ( \vec { x } ) | 3d77393477.png | normalized | |
\widehat { \cal M } _ { i } = K _ { i j } ( \widehat { \Delta } ) s \widehat { \Delta } _ { j } \ \ . | 4068bb05c2.png | tokenized | |
W _ { t r e e } = \lambda \Phi Q { \widetilde Q } ~ , | 104056c5f8.png | normalized | |
t=\int_0^\tau d\tau'\,e(\tau'),\quad T=\int_0^1 d\tau\,e(\tau) | 23e7a0604e.png | none | |
g = e ^ { \frac { i } { 2 } \theta _ { \mathrm { L } } \sigma _ { 2 } } e ^ { \frac { 1 } { 2 } r \sigma _ { 1 } } e ^ { \frac { i } { 2 } \theta _ { \mathrm { R } } \sigma _ { 2 } } , | 761e65e445.png | normalized | |
\eta \cdot k = - i k ^ { 2 } \pm i \sqrt { { \frac { ( k ^ { 2 } + \eta ^ { 2 } ) ( k ^ { 2 } + i \epsilon ) } { 1 - \lambda } } } . | 56dccf2265.png | normalized | |
\label{sequestered}K\ =\ -3\log \left[\, -\frac{1}{3} Q^+ Q \ +\ f(H^+, H)\,\right]\ .\label{K_RS} | 3506377aa4.png | none | |
\vec {k}_1^{ex}= (0,0,0,1):\,\,\,\,\,|C_4|=4, | 7a6324e979.png | none | |
\label{dm}\frac{d G(\rho_2)}{d \rho_2} = - \frac{g_2^2}{8 \pi^2} \rho_2 G(\rho_2)~. | 46a1c22be4.png | none | |
\Omega = D A + A \wedge A \doteq D X ^ { M } \wedge D X ^ { N } \Omega _ { M N } . | 971339cd43.png | normalized | |
K^m_A \frac{\partial}{\partial z_m} = \frac{\partial}{\partial \xi_A} . | 24bea46896.png | none | |
\frac { d ^ { 2 } \xi ^ { \mu } } { d t ^ { 2 } } = - R _ { 0 k 0 } ^ { \mu } \xi ^ { k } - \frac { 1 } { 2 } [ | d754c9bd4a.png | tokenized | |
S_iW_\alpha\bar W_{\dot \alpha},\bar S_iW_\alpha\bar W_{\dot \alpha} | 2378b01281.png | none | |
{ \partial L \over \partial t } = \{ B _ 1 , L \} | 4e70ffd19c.png | tokenized | |
{\cal F}_2 %= - \frac{{\rm ln}Z_2}{\beta V}= \frac{g_r^2M_rT^2}{(4\pi)^2}\left(\frac{1}{x^2}[g_2(x,-r)-g_2(x,r)]^2+[x-h_2(x,-r)-h_2(x,r)]^2\right)\;.\label{F2} | 20c4c431a7.png | none | |
- \partial ^ 2 _ { \rho } \chi + \frac { 2 } { \rho ^ 2 } \chi - \frac { \Lambda ^ 2 } { k ^ 2 } ( 1 - \frac { l ( l + 1 ) k ^ 2 } { E ^ 2 } ) \chi = 0 | 639e7f6e5f.png | tokenized | |
\tilde G ^ { - 1 } ( k ) = \sum _ { s } c ( s ) e ^ { - i k s } | a2f01db623.png | tokenized | |
\label{18}\gamma = \mp f_{\infty} \left( f_{\infty} + {v_0 ^2 \over 2 f_0} \right) | 16ba257815.png | none | |
c _ { 0 j } ^ { N } \ = \ \operatorname { e x p } \left[ \frac { 1 } { 2 } \sum _ { n = 1 } ^ { j } \operatorname { l n } \frac { 1 - n / ( N + 1 ) } { 1 + n / ( N + 1 ) } \right] \ = \ | 4d220a28ce.png | normalized | |
\xi_j = \sqrt{\kappa} z_j\ , \label{wkba} | 425cfe4777.png | none | |
\varepsilon_0 = \langle 0| {\cal H} |0 \rangle = - \sum_{{\bf {p}}, \sigma} \epsilon_{{\bf {p}} \sigma}^{(-)}, | 4990db8e1d.png | none | |
\psi _ { n } ( x ) = x B _ { n - 1 } ( { 1 + \eta \over \sqrt { 1 + 2 \eta } } - x ^ { 2 } ) | 1fbe0941c5.png | tokenized | |
p _ \Lambda = \sum _ { \ell = 1 } ^ { n } c _ { \Lambda \Lambda ^ { ( \ell ) } } \, s _ { \Lambda ^ { ( \ell ) } } , | 5f18fee927.png | tokenized | |
\chi ( T M ) = \chi ( M ) \equiv \sum _ { k } ( - 1 ) ^ { k } b _ { k } ( M ) , | 10e8ca5fd0.png | normalized | |
\Delta _ F ( x , x ' ) = \int \, { d ^ 4 k \over ( 2 \pi ) ^ 4 } { e ^ { i k x } \over k ^ 2 + m ^ 2 - i \epsilon } \, , | 6e9797a29f.png | tokenized | |
\triangle = { \frac { 1 } { \sqrt { g } } } \sum _ { j j ^ { \prime } } { \frac { \partial } { \partial t _ { j } } } \; \left( \, \sqrt { g } \, g ^ { j j ^ { \prime } } \, { \frac { \partial } { \partial t _ { j ^ { \prime } } } } \, \right) | 3d5d30549d.png | normalized | |
\Gamma = A _ { h } \frac { 1 } { ( e ^ { \frac { \omega } { T _ { H } } } - 1 ) } \frac { d ^ 3 k } { ( 2 \pi ) ^ 3 } , | 5050998e2a.png | tokenized | |
{ \cal R } = { \cal R } _ { r } \operatorname { s i n } \eta , ~ ~ ~ \tau = { \cal R } _ { r } ( 1 - \operatorname { c o s } \eta ) . | 1e1bbaef2b.png | normalized | |
{ \widehat M } _ { P } ^ { 2 } \sim { \frac { r ^ { 2 } } { ( \alpha ^ { \prime } ) ^ { 2 } } } ~ , | 58c383b1ae.png | normalized | |
\psi _ { \scriptscriptstyle W } = \frac { 1 } { 2 } ( \psi \pm \gamma _ 5 \psi \gamma _ { 2 1 } ) . | 5aefee37e5.png | tokenized |
End of preview. Expand in Data Studio
IM2LATEX-100K — formula-variant export
IM2LATEX-100K is the image-to-LaTeX dataset introduced by Yuntian Deng, Anssi Kanervisto, and Alexander M. Rush in What You Get Is What You See: A Visual Markup Decompiler (September 2016).
Preferred research citation: Image-to-Markup Generation with Coarse-to-Fine Attention, by Yuntian Deng, Anssi Kanervisto, Jeffrey Ling, and Alexander M. Rush, ICML 2017. This Hugging Face repository is an export maintained by Yuntian Deng.
Canonical project: Image-to-Markup / im2latex · Original downloads and version guide: IM2LATEX-100K data · Original implementation: harvardnlp/im2markup · Training and evaluation guide: Using IM2LATEX-100K
Please cite the original ICML 2017 paper when using IM2LATEX-100K, including this export or a processed copy. For mirrors and derived versions, retain this original citation and credit additional processing or model contributions alongside it. Copyable BibTeX is below.
What “raw” means in this export
This repository contains 203,745 rows: three formula representations for each of the 67,915 source rows represented in the companion normalized export. It was first uploaded to Hugging Face on September 12, 2022. Its source-data revision before this documentation update is 33e16d80cce0f8d043eb8a6f31ede42c10df826b.
raw does not mean the complete original archive or 203,745 independent formula images. It identifies an export with alternative target strings. The preprocessing field is essential:
none: formula source without the tokenization or normalization used for the other target variants.tokenized: formula represented as a space-separated token sequence.normalized: normalized LaTeX token sequence. These(filename, formula)pairs match the companion normalized repository in every split, including repeated rows.
Select the target representation that fits your experiment. If you use multiple representations, report that choice and keep variants of a source example together when making new splits.
Splits and fields
| Split | All variant rows | Rows per preprocessing value | Distinct filenames within split |
|---|---|---|---|
train |
165,099 | 55,033 | 55,021 |
val |
18,216 | 6,072 | 6,072 |
test |
20,430 | 6,810 | 6,810 |
| Total rows | 203,745 | 67,915 | — |
Each split contains equal numbers of none, tokenized, and normalized rows. The validation split is named val, not validation.
image: formula image, stored as an image feature and decoded by Hugging Face Datasets. Thepreprocessinglabel describes the formula target, not an instruction to render or transform this image.formula: LaTeX target string in the representation named bypreprocessing.filename: source image filename. Repeats across representations and is not a unique row identifier.preprocessing: string with one of the valuesnone,tokenized, ornormalized.
Load and choose a target representation
from datasets import load_dataset
dataset = load_dataset("yuntian-deng/im2latex-100k-raw")
raw_targets = dataset.filter(
lambda mode: mode == "none",
input_columns=["preprocessing"],
)
example = raw_targets["train"][0]
image = example["image"]
latex = example["formula"]
Use "tokenized" or "normalized" instead of "none" to select another target representation. To avoid downloading the full export while exploring it, load_dataset(..., streaming=True) also supports filtering. For reproducible experiments, pass revision="<commit hash>". The files are Parquet and require no custom remote loading script.
Origin and scope
The original dataset pairs rendered mathematical expressions with LaTeX source collected from arXiv materials. Its 2016 Zenodo archive records Anssi Kanervisto, a collaborator on the Image-to-Markup project, as the archive creator. The archive points to the original Image-to-Markup code and paper. Preserve both the research authorship and the archive attribution when documenting provenance.
The broader 2016 research studied recovering presentational markup from images, including image-to-LaTeX and image-to-HTML. IM2LATEX-100K is its mathematical-expression dataset. This repository contains formula images and LaTeX targets; it does not contain the project's HTML dataset or handwritten-expression evaluation data.
The name “100K” refers to the original dataset family. This Hugging Face export has its own row counts and preprocessing and must not be described as the complete original archive. Use the canonical data page for the historical downloads and their split definitions.
Relationship to the original files
The exported filenames belong to the corresponding original and filtered splits at the project download page. The normalized targets match im2latex_formulas.norm.lst, and the tokenized targets in this export match im2latex_formulas.tok.lst, using the formula indexes in the source split lists. This export retains a smaller subset of the hosted splits; the exact additional selection and image export recipe is not recorded in the Hugging Face repositories.
The variant export's none targets match im2latex_formulas.lst after decoding UTF-8 while ignoring invalid bytes and stripping surrounding whitespace. In 13 training rows and one validation row, the source contains invalid UTF-8 bytes absent from the export. Use the original file when byte-for-byte source preservation matters.
Intended use and limitations
Use this data for research on mathematical expression recognition, image-to-LaTeX generation, visual sequence modeling, and evaluation of formula transcription. The supervision comes from LaTeX source and rendering, rather than human transcription of photographs.
- The data represents rendered scientific formulas. Performance on it does not establish performance on handwriting, scanned pages, photographs, or complete documents.
- LaTeX has multiple strings with equivalent visual output. Report target normalization, tokenization, rendering setup, and evaluation metric. String accuracy and rendered-image accuracy measure different things.
- Preserve the provided splits for comparisons and report the exact repository revision and any additional filtering. Do not compare this export's results directly with the paper's full test set without aligning the data and protocol.
- The existing normalized export has 12 training filenames occurring twice, four filenames shared by training and validation, and two shared by training and test. The variant export inherits these memberships. A filename is not a globally unique row identifier. These are filename-level checks, not a claim of exhaustive image or semantic deduplication; perform an appropriate contamination audit for your experiment. This documentation update preserves the existing rows.
- The exact 2022 image export/resize recipe is not included in these repositories. Sampled images in all three splits are 320 × 64 pixels; inspect decoded images for your pipeline and use the original project downloads when reproducing its preprocessing. The source project's image preprocessing and formula preprocessing remain available.
License
The original Zenodo dataset record declares CC0 1.0 (cc-zero in the archive metadata). The cc0-1.0 tag on this card records that source dataset license. The original model/code repository separately uses the MIT license. Citation is requested for scholarly attribution; the dataset and code licenses are distinct.
Citation
Download the citation records.
Recommended research and dataset citation:
Yuntian Deng, Anssi Kanervisto, Jeffrey Ling, and Alexander M. Rush. Image-to-Markup Generation with Coarse-to-Fine Attention. Proceedings of the 34th International Conference on Machine Learning, PMLR 70:980–989, 2017. Paper and publisher citation.
@InProceedings{pmlr-v70-deng17a,
title = {Image-to-Markup Generation with Coarse-to-Fine Attention},
author = {Yuntian Deng and Anssi Kanervisto and Jeffrey Ling and Alexander M. Rush},
booktitle = {Proceedings of the 34th International Conference on Machine Learning},
pages = {980--989},
year = {2017},
volume = {70},
series = {Proceedings of Machine Learning Research},
publisher = {PMLR},
url = {https://proceedings.mlr.press/v70/deng17a.html}
}
For the original 2016 Image-to-Markup release and its image-to-HTML work, also cite the dated first version:
@article{deng2016visualmarkup,
title = {What You Get Is What You See: A Visual Markup Decompiler},
author = {Yuntian Deng and Anssi Kanervisto and Alexander M. Rush},
journal = {arXiv preprint arXiv:1609.04938v1},
year = {2016},
url = {https://arxiv.org/abs/1609.04938v1}
}
To identify the source data archive precisely, add its dataset reference:
@dataset{kanervisto2016im2latex,
author = {Anssi Kanervisto},
title = {im2latex-100k, arXiv:1609.04938},
year = {2016},
publisher = {Zenodo},
version = {v1},
doi = {10.5281/zenodo.56198},
url = {https://doi.org/10.5281/zenodo.56198}
}
When redistributing a subset or reformatted copy, link to the canonical project and dataset page, retain the original research citation, and describe your changes. For an experiment using this repository, record its Hugging Face repository name, commit revision, split, and preprocessing choice.
- Downloads last month
- 145