---
license: mit
language:
- fr
task_categories:
- image-to-text
pretty_name: Belfort-line
dataset_info:
  features:
  - name: image
    dtype: image
  - name: text
    dtype: string
  splits:
  - name: train
    num_examples: 25800
  - name: validation
    num_examples: 3102
  - name: test
    num_examples: 3819
  dataset_size: 32721
tags:
- atr
- htr
- ocr
- historical
- handwritten
---

# Belfort - line level

## Table of Contents
- [Belfort - line level](#belfort-line-level)
  - [Table of Contents](#table-of-contents)
  - [Dataset Description](#dataset-description)
    - [Languages](#languages)
  - [Dataset Structure](#dataset-structure)
    - [Data Instances](#data-instances)
    - [Data Fields](#data-fields)

## Dataset Description

- **Homepage:** [Belfort city archives](https://teklia.com/blog/202211-belfort-en/)
- **Source:** [Zenodo](https://zenodo.org/records/8041668)
- **Paper:** [Handwritten Text Recognition from Crowdsourced Annotations](https://doi.org/10.1145/3604951.3605517)
- **Point of Contact:** [TEKLIA](https://teklia.com)

## Dataset Summary 

The Belfort dataset includes minutes of the municipal council of the French city of Belfort. 
Text lines were extracted using an automatic model and may contain segmentation errors. The transcriptions were obtained through a crowdsourcing campaign using the [Callico](https://callico.teklia.com/projects/ce9b42d4-23a8-4381-b5bb-459bedc59165/details/) web plateform.

Note that all images are resized to a fixed height of 128 pixels.

### Languages

All the documents in the dataset are written in French.

## Dataset Structure

### Data Instances

```
{
  'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=4300x128 at 0x1A800E8E190,
  'text': 'les intérêts des 30000 francs jusqu'au moment de la'
}
```

### Data Fields


- `image`: a PIL.Image.Image object containing the image. Note that when accessing the image column (using dataset[0]["image"]), the image file is automatically decoded. Decoding of a large number of image files might take a significant amount of time. Thus it is important to first query the sample index before the "image" column, i.e. dataset[0]["image"] should always be preferred over dataset["image"][0].
- `text`: the label transcription of the image.