Datasets:

ADIP-QCRI
/

ADIP-DocLayNet-English

license: cdla-permissive-1.0
language:
  - en
pretty_name: ADIP-DocLayNet-English
size_categories:
  - 10K<n<100K
task_categories:
  - image-to-text
  - object-detection
tags:
  - document-layout-analysis
  - document-parsing
  - ocr
  - reading-order
  - doclaynet
  - adip
dataset_info:
  features:
    - name: page_id
      dtype: string
    - name: image
      dtype: image
    - name: width
      dtype: int32
    - name: height
      dtype: int32
    - name: doc_category
      dtype: string
    - name: collection
      dtype: string
    - name: original_filename
      dtype: string
    - name: page_no
      dtype: int32
    - name: page_hash
      dtype: string
    - name: layout_prompt
      dtype: string
    - name: layout
      dtype: string
    - name: ocr_format
      dtype: string
    - name: ocr_prompt
      dtype: string
    - name: ocr
      dtype: string
  splits:
    - name: train
      num_examples: 69076
    - name: validation
      num_examples: 6477
    - name: test
      num_examples: 4991
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
      - split: validation
        path: data/validation-*.parquet
      - split: test
        path: data/test-*.parquet

ADIP-DocLayNet-English

80,544 English document pages from DocLayNet v1.2, re-annotated in the ADIP (Arabic Document Intelligence Platform) document-parsing format. Each page has two supervised records:

ADIP layout parsing: the page's document metadata, every element as a hierarchy (page → regions → ADIP element kinds) with boxes on a 0–1000 grid, the reading order, and the relations between elements.
Transcription: the page text in reading order, either as ADIP element JSON (id, kind, pixel box, text) or as plain text.

The element kinds, categories and document types come from the ADIP-Naqsh taxonomy (162 element kinds in 9 categories). The records use the same fields and value conventions as ADIP's Arabic document annotations, so English and Arabic pages can be mixed in one training set or evaluated with one parser.

Split	Pages	ADIP layout records	Transcription records (JSON / plain text)
train	69,076	69,076	34,138 / 34,554
validation	6,477	6,477	3,186 / 3,260
test	4,991	4,991	2,493 / 2,458

The splits are DocLayNet's own. A page whose box has no text (for example, a page that is one picture) has a layout record but no transcription. 41 pages whose layout record exceeded about 12k tokens were left out.

Loading

import json
from datasets import load_dataset

ds = load_dataset("ADIP-QCRI/ADIP-DocLayNet-English", split="validation")
row = ds[0]
row["image"]                    # PIL image of the page (JPEG, 1025 x 1025)
layout = json.loads(row["layout"])
row["ocr_format"], row["ocr"]   # "json" or "text"

Columns

Column	Content
`page_id`	`doclaynet_<first 16 hex digits of the DocLayNet page hash>`
`image`	the page image, as in DocLayNet (1025 × 1025 px)
`width`, `height`	image size in pixels
`doc_category`, `collection`, `original_filename`, `page_no`, `page_hash`	DocLayNet source metadata
`layout_prompt`, `layout`	the ADIP layout-parsing instruction and its answer (compact JSON string)
`ocr_format`, `ocr_prompt`, `ocr`	`json` or `text`, the transcription instruction and its answer

The ADIP layout record

{
  "document": {
    "type": "financial_report",
    "script": {"primary": "en", "direction": "ltr", "diacritics": "none", "ligatures": "none",
               "numerals": "western", "kashida": "none"},
    "script_style": "system_serif_latin", "estimated_century": "21st",
    "paper_color": "white", "ink_color": "black", "readability": 5, "complexity": 3,
    "page_size_px": [1025, 1025]
  },
  "reading_order": ["elem_0001_01", "elem_0001_02"],
  "reading_order_elements": ["elem_0001", "elem_0001_01", "elem_0001_01_01", "..."],
  "elements": [{
    "id": "elem_0001", "slot": "page", "kind": "page", "bbox_2d": [0, 0, 1000, 1000], "level": 0, "...": "...",
    "children": [{
      "id": "elem_0001_01", "slot": "main_text_region", "kind": "body", "bbox_2d": [116, 115, 942, 743],
      "level": 1, "parent": "elem_0001", "script": "en", "direction": "ltr", "layer": 1,
      "filled": true, "fill_ratio": 0.97, "order": 2,
      "children": [{
        "id": "elem_0001_01_01", "kind": "body_text", "category": "structural_body",
        "bbox_2d": [147, 88, 837, 118], "level": 2, "parent": "elem_0001_01",
        "font": "system_serif_latin", "ink": "black", "script": "en", "direction": "ltr",
        "layer": 2, "lines": 2, "order": 3
      }]
    }]
  }],
  "relations": [{"type": "next", "from": "elem_0001_01_01", "to": "elem_0001_01_02"}]
}

document: ADIP document type, script profile, script style, estimated century, paper and ink colour, readability (1–5), complexity (1–5) and page size.
elements: the page; its regions (header_region, main_text_region, footer_region); and the ADIP elements inside them. Each element has its id, slot or kind, ADIP category, bbox_2d = [x0, y0, x1, y1] on a 0–1000 grid of the page, level, parent, font class, ink colour, script, direction, layer, number of text lines and position in the reading order.
reading_order lists the region ids in reading order. reading_order_elements lists every element id in reading order.
relations: next between consecutive elements in reading order, and inside for an element whose box lies at least 90% inside a larger element of another region.

The transcription record

json: [{"id", "kind", "bbox": {"x", "y", "w", "h"} in pixels, "text"}, ...], one entry per text-bearing element, in reading order. The text comes from the PDF text cells of the element's box, grouped into lines.
text: the same text as plain text, one element per paragraph, in reading order.

DocLayNet classes → ADIP element kinds

DocLayNet	ADIP kind	ADIP category	Region
Text	body_text	structural_body	main_text_region
Section-header	heading	structural_body	main_text_region
List-item	list	structural_body	main_text_region
Caption	caption	structural_body	main_text_region
Title	title	metadata	main_text_region
Table	table	block_assets	main_text_region
Picture	figure	block_assets	main_text_region
Formula	math_block	block_assets	main_text_region
Page-header	header	navigation_reference	header_region
Page-footer	page_number if the text is a page number, else footer	navigation_reference	footer_region
Footnote	footnote	navigation_reference	footer_region

DocLayNet document categories map to ADIP document types: financial reports → financial_report, manuals → technical_manual, scientific articles → scientific_article, laws and regulations → regulation, patents → patent, government tenders → specification.

Element counts over all splits: body_text 510,377; list 185,660; heading 142,884; header 58,022; figure 45,976; page_number 39,758; table 34,733; footer 31,120; math_block 25,027; caption 22,524; footnote 6,318; title 5,071.

Derived values

Boxes, classes and text are DocLayNet's own annotations. DocLayNet has no reading order and no visual metadata, so these values are derived by fixed rules. Treat them as estimates:

Reading order: the header region left to right, then the main region, then the footer region (footnotes first, then page footers). In the main region, a block wider than 60% of the text area is read on its own; the blocks between such blocks are read column by column, with the column taken from the box centre.
Fonts (script_style, element font): system_serif_latin, system_sans_latin or system_mono_latin, from the PDF font names. ADIP's font taxonomy covers Arabic script, and these are the Latin counterparts of its system_serif_arabic and system_sans_arabic classes.
Colours: paper_color comes from the page background pixels. ink_color and element ink come from the colour of the PDF text cells.
estimated_century: from a year in the source file name, otherwise 21st.
readability is 5 (born-digital pages). complexity (1–5) comes from the number of blocks, the number of columns and the number of tables, figures and formulas.
Lines: the number of text lines in the element, from its PDF text cells.

License and attribution

This dataset is a modified version of DocLayNet v1.2 and is distributed under the same license, CDLA-Permissive-1.0 (full text in LICENSE). The page images, boxes, classes, text cells and source metadata are © the DocLayNet authors and data providers. The modifications are the ADIP re-annotation described above. If you use this dataset, please cite DocLayNet:

@inproceedings{pfitzmann2022doclaynet,
  title     = {DocLayNet: A Large Human-Annotated Dataset for Document-Layout Segmentation},
  author    = {Pfitzmann, Birgit and Auer, Christoph and Dolfi, Michele and Nassar, Ahmed S. and Staar, Peter},
  booktitle = {Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  pages     = {3743--3751},
  year      = {2022},
  doi       = {10.1145/3534678.3539043}
}

Contact

Ehsan Asgari, easgari@hbku.edu.qa