File size: 7,245 Bytes
4661efe a76d80f 4661efe a76d80f 4661efe 6427254 4661efe 6427254 4661efe a76d80f 4661efe a76d80f 4661efe a76d80f 4661efe a76d80f 4661efe a76d80f 4661efe a76d80f 4661efe a76d80f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 | ---
pretty_name: PDFluent PDF Document Layout Ground Truth (sample)
license: other
license_name: mixed-per-config
license_link: https://huggingface.co/datasets/PDFluent/pdf-document-layout-ground-truth#licences-and-attribution
language:
- en
size_categories:
- 1K<n<10K
task_categories:
- object-detection
- image-to-text
tags:
- pdf
- document-ai
- document-understanding
- layout-analysis
- table-extraction
- heading-detection
- math-formula-recognition
- latex
- ocr
- rag
- ground-truth
- pdf-to-markdown
- document-parsing
configs:
- config_name: default
default: true
data_files:
- split: train
path: data/default/train.jsonl
- config_name: third_party_scans
data_files:
- split: train
path: data/third_party_scans/train.jsonl
- config_name: share_alike
data_files:
- split: train
path: data/share_alike/train.jsonl
---
# PDFluent PDF Document Layout Ground Truth Dataset
Block-level layout annotations for born-digital PDFs (headings, paragraphs, tables, reading order) with LaTeX formulas and scan pages. For researchers and developers building document parsing, RAG tools, and LLM pipelines. Made by the team behind PDFluent SDK 1.3.0. PDFluent 1.3.0 ranks above every engine on the published opendataloader-bench leaderboard (as of 3 November 2026)*.
\*Overall score 0.9168 (quality="auto", keep_furniture=True, Linux) against the engines listed in the benchmark README on 3 November 2026; submitted as pull request #37 (https://github.com/opendataloader-project/opendataloader-bench/pull/37). Other engines have submissions open that are not yet on the leaderboard.
## Use it in two lines
```python
from datasets import load_dataset
pages = load_dataset("PDFluent/pdf-document-layout-ground-truth", "default", split="train")
```
The PDF behind each row is in this repository at the path in the `asset` field.
## More examples
```python
import json
from datasets import load_dataset
repo = "PDFluent/pdf-document-layout-ground-truth"
pages = load_dataset(repo, "default", split="train").filter(lambda r: r["kind"] == "page")
formulas = load_dataset(repo, "default", split="train").filter(lambda r: r["kind"] == "formula")
scans = load_dataset(repo, "third_party_scans", split="train")
sa = load_dataset(repo, "share_alike", split="train")
row = pages[0]
annotation = json.loads(row["annotation_json"])
print(row["document_id"], row["page_number"], list(annotation))
```
The PDF behind each row is in this repository at the path in `asset`:
```python
from huggingface_hub import hf_hub_download
pdf_path = hf_hub_download(repo, row["asset"], repo_type="dataset")
```
## What you get
- CC BY 4.0 annotations for born-digital PDFs (free for commercial use with attribution)
- FinTabNet scans under CDLA-Permissive-1.0 terms
- CC BY-SA 4.0 config for share-alike projects
- Evaluate layout analysis, table extraction, and heading detection models
- Lightly fine-tune formula recognition and reading order systems
## What is in it
The dataset has three configs. `default`: born-digital PDF pages with block-level layout annotations (headings, paragraphs, tables) and display-formula LaTeX. `third_party_scans`: scan pages from FinTabNet, with IBM's ground truth. `share_alike`: pages from CC BY-SA papers, with annotations under CC BY-SA.
## Fields
Each row links to a PDF page asset by SHA-256. annotation_json is a string because page, scan and formula ground truth schemas differ.
| Field | Meaning |
|---|---|
| `document_id` | Source document identifier (for example `arxiv:...`) |
| `source` | Source collection |
| `kind` | `page`, `formula` or `scan` |
| `page_number` | Page in the source document |
| `asset`, `asset_sha256` | Path of the PDF file in this repo and its SHA-256 |
| `annotation_json` | Ground truth as a JSON string (blocks, boxes, reading order, source text, LaTeX) |
| `annotation_licence`, `annotation_attribution` | Licence and credit for the annotation |
| `paper_licence`, `paper_attribution_json` | Licence and attribution of the underlying paper |
There is one split, `train`. There is no evaluation split.
## Sampling
This is a sample, not the full ground truth. Documents were selected at document level, stratified by source. The sample supports evaluation and light fine-tuning but is too small to train a model to PDFluent's level.
## Holdouts and label quality
Documents in PDFluent's benchmark and holdout sets are excluded. The exclusion registry has partial coverage for some sources. Labels are generated automatically and inherit extraction errors.
## Licences and attribution
Our annotations in `default` are CC BY 4.0: credit 'Innovation Trigger B.V., PDFluent ground-truth annotations (2026)' and link to https://pdfluent.com. Paper content keeps its original licence. `third_party_scans` keeps CDLA-Permissive-1.0. `share_alike` annotations are CC BY-SA 4.0.
| Config | Annotation licence | Underlying content |
|---|---|---|
| `default` | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) | arXiv (CC BY 4.0, CC0 1.0), PubMed Central (CC BY), synthetic pages |
| `third_party_scans` | [CDLA-Permissive-1.0](https://cdla.dev/permissive-1-0/) | FinTabNet, IBM Research |
| `share_alike` | [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/) | arXiv papers under CC BY-SA 4.0 |
Mandatory credit for `default`: **Innovation Trigger B.V., PDFluent ground-truth annotations (2026)**, https://pdfluent.com. Keep the credit, the licence link and an indication of your changes. Per-document paper credit is in [ATTRIBUTION.tsv](ATTRIBUTION.tsv). The package as a whole is mixed-source and has no blanket licence; check the config you use.
## Licence in plain words
Default config uses our CC BY 4.0 annotations (credit line: Innovation Trigger B.V., PDFluent ground-truth annotations (2026)). Third-party scans follow CDLA-Permissive-1.0 terms. Share-alike config requires CC BY-SA 4.0 for derivatives.
## Try PDFluent
Load this dataset in two lines of code. Run the PDFluent SDK on the same PDFs and diff its Markdown against the annotations. See pdfluent.com/docs.
## Citation
```bibtex
@misc{pdfluent_layout_ground_truth_2026,
author = {{Innovation Trigger B.V.}},
title = {PDFluent PDF document layout ground truth (document-level sample)},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/PDFluent/pdf-document-layout-ground-truth},
note = {PDFluent, https://pdfluent.com}
}
```
## Links
- Website: https://pdfluent.com
- Documentation: https://pdfluent.com/docs
- SDK on PyPI: https://pypi.org/project/pdfluent/
- SDK source and releases: https://github.com/pdfluent/pdfluent-sdk
- Licence: https://pdfluent.com/license
- Organisation: https://huggingface.co/PDFluent
- Formula recognition model: https://huggingface.co/PDFluent/formula-recognition-onnx
- Heading level classifier: https://huggingface.co/PDFluent/pdf-heading-level-classifier
- Layout ground-truth dataset: https://huggingface.co/datasets/PDFluent/pdf-document-layout-ground-truth
- Collection: https://huggingface.co/collections/PDFluent/pdfluent-document-ai-6ae99736adab8c6d90781ab5
|