---
license: other
pretty_name: OpenPSS (community mirror)
tags:
- page-stream-segmentation
- document-ai
- document-boundary-detection
- benchmark
- mirror
task_categories:
- image-classification
dataset_info:
- config_name: LONG
  features:
  - name: stream_id
    dtype: string
  - name: position
    dtype: int32
  - name: image
    dtype: image
  - name: text
    dtype: string
  - name: label
    dtype: int8
  splits:
  - name: train
    num_bytes: 998698445
    num_examples: 63815
  - name: test
    num_bytes: 405971545
    num_examples: 25676
  download_size: 1334166455
  dataset_size: 1404669990
- config_name: SHORT
  features:
  - name: stream_id
    dtype: string
  - name: position
    dtype: int32
  - name: image
    dtype: image
  - name: text
    dtype: string
  - name: label
    dtype: int8
  splits:
  - name: train
    num_bytes: 976671526
    num_examples: 40715
  - name: test
    num_bytes: 293561622
    num_examples: 11462
  download_size: 1224066290
  dataset_size: 1270233148
configs:
- config_name: LONG
  data_files:
  - split: train
    path: LONG/train-*
  - split: test
    path: LONG/test-*
- config_name: SHORT
  data_files:
  - split: train
    path: SHORT/train-*
  - split: test
    path: SHORT/test-*
---



# OpenPSS — community mirror

> ⚠️ **This is a redistribution (mirror) of the OpenPSS benchmark, not our own work.** It is hosted for
> availability and reproducibility. **All credit belongs to the original authors.** If you are an author or
> rights-holder and would like any change or removal, please open a discussion here or contact us.

## Original work
**OpenPSS: An Open Page Stream Segmentation Benchmark** — Ruben van Heusden, Jaap Kamps, Maarten Marx
(University of Amsterdam, IRLab), *TPDL 2024* (DOI [10.1007/978-3-031-72437-4_24](https://link.springer.com/chapter/10.1007/978-3-031-72437-4_24)).
Builds on **WooIR** (van Heusden, Kamps, Marx, *SIGIR ICTIR 2022*).

## What this is
**Page Stream Segmentation (PSS):** given a stream of concatenated scanned pages, recover the original
document boundaries. OpenPSS provides two datasets — **SHORT** and **LONG**. The source documents are
Dutch government records released under the Freedom of Information Act (*Wet open overheid* / Woo), which
are public records.

This is a **self-contained, HuggingFace-native reformatting** of the original OpenPSS release (page images
embedded — no external `png.zip` to unpack), exposed as two loadable **configs** with train/test splits:

```python
from datasets import load_dataset
short = load_dataset("nutrientdocs/openpss-mirror", "SHORT")   # or "LONG"
# columns: stream_id (str), position (int, 1-indexed page in the stream),
#          image (PIL), text (OCR), label (int: 1 = starts a new document / boundary, else 0)
```
A stream's ground-truth boundaries are the `label` column ordered by `position`; `label[0]` (page 1) is 1
by construction. Content is faithful to the original OpenPSS SHORT/LONG; only the packaging is changed.

## Citation
```bibtex
@inproceedings{vanHeusden2024OpenPSS,
  title     = {OpenPSS: An Open Page Stream Segmentation Benchmark},
  author    = {van Heusden, Ruben and Kamps, Jaap and Marx, Maarten},
  booktitle = {Theory and Practice of Digital Libraries (TPDL)},
  year      = {2024}, doi = {10.1007/978-3-031-72437-4_24}
}
```

## License / terms
The underlying documents are Dutch FOIA (Woo) **public records**; OpenPSS is distributed as an open
benchmark. This mirror redistributes the material under those terms — please consult the original
publication for authoritative licensing. Attribution to the original authors is required.

*Mirror maintained by Nutrient (`nutrientdocs`) to keep the OpenPSS benchmark reliably available and to make
our page-stream-segmentation results reproducible.*