rootsautomation/TABMEpp

---
language:
- en
license: mit
size_categories:
- 100K<n<1M
task_categories:
- text-classification
- image-classification
pretty_name: TABME++
dataset_info:
features:
- name: doc_id
dtype: string
- name: pg_id
dtype: int32
- name: ocr
dtype: string
- name: img
dtype: binary
splits:
- name: train
num_bytes: 14584489453
num_examples: 110132
- name: val
num_bytes: 796081484
num_examples: 6089
- name: test
num_bytes: 812403045
num_examples: 6237
download_size: 11207258731
dataset_size: 16192973982
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: val
path: data/val-*
- split: test
path: data/test-*
---

# Dataset Card for TABME++

The TABME dataset is a synthetic collection of business document folders generated from the Truth Tobacco Industry Documents archive, with preprocessing and OCR results included, designed to simulate real-world digitization tasks.

TABME++ extends TABME by enriching it with commercial-quality OCR (Microsoft OCR).

## Dataset Details

### Dataset Description

The TABME dataset is a synthetic collection created to simulate the digitization of business documents, derived from a portion of the Truth Tobacco Industry Documents (TTID) archive. The dataset was constructed by sampling 44,769 PDF documents, excluding corrupted files and those longer than 20 pages, and then preprocessing them by cropping margins, converting them to grayscale, and resizing to 1,000 pixels. To mimic real-world scenarios, folders of documents were generated using a Poisson distribution with 𝜆 = 11, leading to a mean folder length of around 30 pages. The dataset was split into training, validation, and test sets, with OCR preprocessing performed using the Tesseract engine. The dataset includes 100,000 folders for training, 5,000 for validation, and 5,000 for testing, and the results include recognized words, their coordinates, and confidence levels.

TABME++ replaces the previous OCR with commericial-quality OCR obtained through Microsoft's OCR services.

- **Curated by:** UCSF, UCL, University of Cambridge, Vector.ai, Roots Automation
- **Language(s) (NLP):** English
- **License:** MIT

### Dataset Sources

- **Repository:**
- [Original TABME release](https://github.com/aldolipani/TABME)
- **Paper:**
- [Tab this Folder of Documents: Page Stream Segmentation of Business Documents](https://dl.acm.org/doi/10.1145/3558100.3563852)

## Uses

### Direct Use

This dataset is intended to be used for page stream segmentation: the segmentation of a stream of ordered pages into coherent atomic documents.

## Dataset Structure

Each row of the dataset corresponds to one page of one document.
Each page has the following features:
- `doc_id`, str: The unique document id this page belongs to
- `pg_id`, int: The page id within its document
- `ocr`, str: A string containing the OCR annotations from Microsoft OCR. These can be loaded as a Python dictionary with `json.loads` (or equivalent).
- `img`, binary: The raw bytes of the page image. These can be converted back to a PIL.Image with `Image.open(io.BytesIO(bytes))` (or equivalent).

This dataset is given such that each document appears once.
To build out the full aggregated synthetic streams, one needs to collate the unique documents according to the streams described in the [streams sub-folder](https://huggingface.co/datasets/rootsautomation/TABMEpp/tree/main/streams).

## Dataset Creation

### Curation Rationale

The original data, Truth Tobacco Industry Documents archive (formerly known as Legacy Tobacco Documents Library), was curated by researchers at UCSF.
This was to promote the study of information retrieval and analysis of business documents.

TABME was curated to promote research on page stream segmentation, a core task in automated document processing.
TABME++ improves upon TABME by adding higher-quality OCR annotations, but is still curated for the same purposes.

### Source Data

From the [UCSF Library](https://www.industrydocuments.ucsf.edu/tobacco/):
> Truth Tobacco Industry Documents (formerly known as Legacy Tobacco Documents Library) was created in 2002 by the UCSF Library. It was built to house and provide permanent access to tobacco industry internal corporate documents produced during litigation between US States and the seven major tobacco industry organizations and other sources. These internal documents give a view into the workings of one of the largest and most influential industries in the United States.

## Citation

**BibTeX:**

```
@misc{heidenreich2024largelanguagemodelspage,
title={Large Language Models for Page Stream Segmentation},
author={Hunter Heidenreich and Ratish Dalvi and Rohith Mukku and Nikhil Verma and Neven Pičuljan},
year={2024},
eprint={2408.11981},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2408.11981},
}
```

## Dataset Card Authors

Hunter Heidenreich