metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: unique_id
      dtype: string
    - name: text
      dtype: string
    - name: text_line_id
      dtype: string
    - name: raw_text
      dtype: string
    - name: text_transform
      dtype: string
    - name: font_path
      dtype: string
    - name: font_size
      dtype: int64
    - name: text_color
      dtype: string
    - name: background_color
      dtype: string
    - name: top_margin
      dtype: int64
    - name: bottom_margin
      dtype: int64
    - name: left_margin
      dtype: int64
    - name: right_margin
      dtype: int64
    - name: bbox_left
      dtype: int64
    - name: bbox_top
      dtype: int64
    - name: bbox_right
      dtype: int64
    - name: bbox_bottom
      dtype: int64
    - name: image_width
      dtype: int64
    - name: image_height
      dtype: int64
    - name: undistorted_file_name
      dtype: string
    - name: augraphy_log_path
      dtype: string
    - name: distorted_bbox_left
      dtype: float64
    - name: distorted_bbox_top
      dtype: float64
    - name: distorted_bbox_right
      dtype: float64
    - name: distorted_bbox_bottom
      dtype: float64
    - name: language_code
      dtype: string
  splits:
    - name: train
      num_bytes: 10298088981.391
      num_examples: 307387
    - name: validation
      num_bytes: 1353982194.08
      num_examples: 40765
    - name: test
      num_bytes: 2455873836.834
      num_examples: 84534
  download_size: 13836005253
  dataset_size: 14107945012.305
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
license: cc-by-3.0
task_categories:
  - image-to-text
tags:
  - OCR
  - ATR
  - Sámi
  - synthetic
pretty_name: Synthetic OCR data for North, South, Lule and Inari Sámi
size_categories:
  - 100K<n<1M
language:
  - sma
  - sme
  - smj
  - smn
  - smi
  - se

Synthetic text images for North, South, Lule and Inari Sámi

This dataset contains synthetic line images meant for fitting OCR models for North, South, Lule and Inari Sámi. Clean line images are created using Pillow and they are subsequently distorted using Augraphy [1].

Text sources

The text in this dataset comes from Giellatekno's corpus. Specifically, we used the data files of the converted/-directories of [2][3][4][5] (commit hashes 32f4af263cefae6ab9182638e2451ff151757adc, 00dac0e9e74b4a89214ad7d34de27b83362b3f3a, 4303edf80ae5eee2a036663c7b38756a0aa2a189, 7e3437ce8c7dc7692ccbd2505412c03e9e617be6).

Splits

The dataset is split randomly by file so 71 % of the files (307387 lines) are in the training split, 9 % of the files (40765 lines) are in the validation split and 20 % of the files (84534 lines) are in the test split. Each split has a unique set of typefaces and text/background colors.

Language distribution

The language distribution for the different languages are

Language Code	Num train lines	Num val lines	Num test lines
sma	76971	10992	21981
sme	76949	10992	21990
smj	76970	9081	20465
smn	76497	9700	20098

Code

The code to create this dataset is available on our GitHub Repo (commit hash 90341bc19d6368c7848dcc2459065058486a89ea).

Referencing the dataset

If you use this dataset in your research, then please cite both "Enstad T, Trosterud T, Røsok MI, Beyer Y, Roald M. Comparative analysis of optical character recognition methods for Sámi texts from the National Library of Norway. Accepted for publication in Proceedings of the 25th Nordic Conference on Computational Linguistics (NoDaLiDa) 2025." (see the paper repository) and the SIKOR dataset the Sámi text is from: "SIKOR UiT The Arctic University of Norway and the Norwegian Saami Parliament's Saami text collection, http://gtweb.uit.no/korp, Version 01.12.2021 [Data set]." Also note that the SIKOR dataset to get Sámi text for the images is (CC-BY 3.0) licensed.

Dataset license

The dataset is licensed with a CC-BY 3.0 license.