IM2LATEX-100K

Origin and attribution.

IM2LATEX-100K was introduced in Deng, Kanervisto, and Rush’s 2016 visual markup decompiler paper and described in the 2017 ICML paper by Deng, Kanervisto, Ling, and Rush. The original formulas were extracted from arXiv LaTeX sources in the Cornell KDD Cup collection and rendered as images.

The original Zenodo archive was prepared and deposited by coauthor Anssi Kanervisto; its displayed publication date is 21 June 2016. The project’s code repository provides preprocessing, training, and evaluation.

Choose the right version

One lineage, distinct distributions.

The paper’s benchmark description, hosted file lists, and later Hugging Face exports have different counts. These are documented separately so experiments can identify their exact input data.

The hosted formula lists each contain 103,559 indexed lines. Preserve empty entries in the tokenized and normalized lists: removing them changes formula indices. The hosted counts describe the files served here; they do not revise the paper’s reported experiment sizes. File sizes and hashes · SHA-256 checksums

Direct downloads

Original project files.

The existing file addresses remain unchanged. Use the original repository’s preprocessing and evaluation scripts when reproducing its pipeline.

Processed files

The preprocessing pipeline crops images, downsamples and pads them for batching, and tokenizes or normalizes formula targets. Training and validation are filtered; the test list is converted to the processed format without filtering.

• im2latex_formulas.norm.lstNormalized LaTeX targets · 103,559 indexed lines
• im2latex_formulas.tok.lstTokenized LaTeX targets · 103,559 indexed lines
• formula_images_processed.tar.gzCropped, downsampled and padded expression images
• im2latex_train_filter.lstProcessed training list · 75,275 records
• im2latex_validate_filter.lstProcessed validation list · 8,370 records
• im2latex_test_filter.lstFormat-converted test list · 10,355 records; not filtered

Raw files

Formula source and original split lists. Raw list records use formula_index image_name render_type; processed records use image_filename formula_index. Formula indices are zero-based.

• im2latex_formulas.lstOriginal LaTeX targets · 103,559 indexed lines
• im2latex_train.lstRaw training list · 83,884 records
• im2latex_validate.lstRaw validation list · 9,320 records
• im2latex_test.lstRaw test list · 10,355 records
• formula_images.tar.gzOriginal rendered images

IM2LATEX-100K-HANDWRITTEN

The synthetic handwritten-expression release used for pretraining in the paper, distinct from the printed-expression dataset.

• IM2LATEX-100K-HANDWRITTEN.tgzSynthetic handwritten images, formula targets and split files

Hugging Face

Ready-to-load exports.

from datasets import load_dataset

# Normalized targets; split names are train, val, test.
data = load_dataset("yuntian-deng/im2latex-100k")

# Three target representations, labeled by preprocessing.
variants = load_dataset("yuntian-deng/im2latex-100k-raw")

These exports are subsets of the hosted project files, not interchangeable copies of the full benchmark. Consult their dataset cards for exact provenance, repeated filenames, and evaluation limitations; record the repository revision in your experiments.

Reproducibility

Before training or comparing results.