---
license: other
task_categories:
- object-detection
- image-to-text
language:
- uk
- ru
tags:
- handwriting
- document-layout
- math
- coco
- mathglyph-pages
configs:
- config_name: default
  data_files:
  - split: train
    path:
    - train/images/*.png
    - train/metadata.jsonl
  - split: validation
    path:
    - validation/images/*.png
    - validation/metadata.jsonl
---

# MathGlyph Pages 1k With Detector Boxes

Synthetic mixed handwritten math pages prepared for the DFS/Rukopys detector pretraining pipeline.

## Links

- Generator repo: [reirei-00/mathglyph_pages](https://github.com/reirei-00/mathglyph_pages)

## Format

- `train/images/`: train page images.
- `validation/images/`: validation page images.
- `annotations/instance_train.json`: COCO-style detector annotations for train.
- `annotations/instance_val.json`: COCO-style detector annotations for validation.
- `train/metadata.jsonl`: row-level train metadata for Hugging Face image-folder loading.
- `validation/metadata.jsonl`: row-level validation metadata for Hugging Face image-folder loading.
- `classes.txt`: detector class order.
- `summary.json`: split and class counts.

## Classes

`handwritten`, `printed`, `formula`, `table`, `annotation`, `image`, `graph`

## Counts

Train: 900 pages, 21,634 boxes.
Validation: 100 pages, 2,346 boxes.

This dataset is intended as weak/synthetic pretraining data.