harryrobert/latex-ocr

LaTeX OCR Dataset

A dataset for training LaTeX OCR models that convert images of mathematical formulas into LaTeX source code. Built by merging and re-splitting multiple public sources, then applying two levels of augmentation for two-stage training.

---

Dataset Summary

Property	Value
Total unique samples	~732,952
Train (stage1)	659,658
Train (stage2)	659,658
Validation	36,647
Test	36,647
Image height	64 px (fixed)
Image width	16 – 672 px (variable, aligned to 16px)
Label format	Raw LaTeX string
Max token length	200 tokens

---

Sources

All data is merged from the following public datasets, filtered, shuffled, then re-split 90/5/5:

Dataset	Config	Splits used
linxy/LaTeX_OCR	full	train + validation + test
linxy/LaTeX_OCR	synthetic_handwrite	train + validation + test
linxy/LaTeX_OCR	human_handwrite	train + validation + test
OleehyO/latex-formulas	cleaned_formulas	train

Filtering: samples with fewer than 2 or more than 200 LaTeX tokens are removed.

---

Splits

stage1 — Light augmentation (for Stage 1 training: encoder warm-up)

Same 659,658 training samples as stage2, but with lighter augmentation. Each image is independently re-augmented, so stage1 and stage2 are not identical.

Augmentations applied (each with independent probability):

Augmentation	Probability	Parameters
Gaussian blur	0.30	radius 0.3 – 1.2
Rotation	0.30	–3° to +3°
Background color blend	0.40	random warm background
Edge shadow	0.20	left / right / top / bottom
Low resolution	0.20	downscale 20–60% then upscale

35% of samples are kept clean (no augmentation applied).

---

stage2 — Heavy augmentation (for Stage 2 training: LoRA fine-tuning)

Same training indices as stage1, re-augmented with a heavier pipeline.

Augmentations applied (each with independent probability):

Augmentation	Probability	Parameters
JPEG compression	0.40	quality 30 – 75
Low resolution	0.40	downscale 20–60% then upscale
Gaussian noise	0.35	std 5 – 25
Salt & pepper noise	0.20	amount 1 – 5%
Gaussian blur	0.30	radius 0.3 – 1.2
Background color blend	0.40	random warm background
Rotation	0.35	–3° to +3°
Perspective distortion	0.25	deviation 2 – 6%
Random erase	0.30	1–3 rectangles, white/black/gray fill
Edge shadow	0.25	left / right / top / bottom

35% of samples are kept clean (no augmentation applied).

---

validation and test

Drawn from the same shuffled pool (5% each). No augmentation — images are only resized to 64px height.

---

Data Format

Each row contains two columns:

Column	Type	Description
image	PIL.Image (JPEG)	Formula image, height=64px, width≤672px
label	str	Ground-truth LaTeX string

Image preprocessing:

• Convert to RGB
• Resize to height=64px (width scaled proportionally)
• Width clamped to 672px maximum
• Width aligned to nearest multiple of 16px (patch size)

LaTeX tokenization (for reference, not stored):

import re
tokens = re.findall(r"\\[a-zA-Z]+|[^\s]", label)

---

Usage

Load the full dataset

from datasets import load_dataset

ds = load_dataset("harryrobert/latex-ocr")
print(ds)
# DatasetDict({
#     stage1: Dataset({features: ['image', 'label'], num_rows: 659658}),
#     stage2: Dataset({features: ['image', 'label'], num_rows: 659658}),
#     validation: Dataset({features: ['image', 'label'], num_rows: 36647}),
#     test: Dataset({features: ['image', 'label'], num_rows: 36647})
# })

Load a specific split

train_stage1 = load_dataset("harryrobert/latex-ocr", split="stage1")
val = load_dataset("harryrobert/latex-ocr", split="validation")

sample = train_stage1[0]
print(sample["label"])   # e.g. '\frac{1}{2}'
sample["image"].show()   # PIL Image

Streaming (large splits)

ds = load_dataset("harryrobert/latex-ocr", split="stage1", streaming=True)
for sample in ds.take(5):
    print(sample["label"])

PyTorch DataLoader integration

from datasets import load_dataset
from torch.utils.data import DataLoader

ds = load_dataset("harryrobert/latex-ocr", split="stage1")
ds = ds.with_format("torch")

loader = DataLoader(ds, batch_size=32, shuffle=True)

---

Training Recipe

This dataset is designed for a two-stage training pipeline:

Stage 1 — Train visual encoder, freeze language model decoder:

train on split="stage1"
evaluate on split="validation"

Stage 2 — LoRA fine-tuning of the full model:

train on split="stage2"
evaluate on split="validation"

Final evaluation:

evaluate on split="test"

---

License

Dataset contents are derived from linxy/LaTeX_OCR and OleehyO/latex-formulas. Please refer to the original datasets for their respective licenses.