Improve dataset card

README.md

@@ -21,287 +21,61 @@ task_ids: []
 
 # YALTAi Tabular Dataset
 
-## Table of Contents
-- [YALTAi Tabular Dataset](#YALTAi-Tabular-Dataset)
-  - [Table of Contents](#table-of-contents)
-  - [Dataset Description](#dataset-description)
-    - [Dataset Summary](#dataset-summary)
-    - [Supported Tasks and Leaderboards](#supported-tasks-and-leaderboards)
-  - [Dataset Structure](#dataset-structure)
-    - [Data Instances](#data-instances)
-    - [Data Fields](#data-fields)
-    - [Data Splits](#data-splits)
-  - [Dataset Creation](#dataset-creation)
-    - [Curation Rationale](#curation-rationale)
-    - [Source Data](#source-data)
-      - [Initial Data Collection and Normalization](#initial-data-collection-and-normalization)
-      - [Who are the source language producers?](#who-are-the-source-language-producers)
-    - [Annotations](#annotations)
-      - [Annotation process](#annotation-process)
-      - [Who are the annotators?](#who-are-the-annotators)
-    - [Personal and Sensitive Information](#personal-and-sensitive-information)
-  - [Considerations for Using the Data](#considerations-for-using-the-data)
-    - [Social Impact of Dataset](#social-impact-of-dataset)
-    - [Discussion of Biases](#discussion-of-biases)
-    - [Other Known Limitations](#other-known-limitations)
-  - [Additional Information](#additional-information)
-    - [Dataset Curators](#dataset-curators)
-    - [Licensing Information](#licensing-information)
-    - [Citation Information](#citation-information)
-    - [Contributions](#contributions)
+353 page images of historical documents with tabular layouts, mostly notarial registers, with bounding-box annotations for four zone types: `Col`, `Header`, `Marginal` and `text`. Created by Thibault Clérice and deposited on [Zenodo](https://doi.org/10.5281/zenodo.6827706) alongside the paper [You Actually Look Twice At it (YALTAi)](https://arxiv.org/abs/2207.11230) (*Journal of Data Mining and Digital Humanities*, 2022), which treats page layout recognition on historical documents as an object detection problem rather than a pixel classification one.
 
-## Dataset Description
+The annotations support layout analysis before transcription. Telling the columns of a register apart from its header and its marginalia is what decides whether a downstream OCR/HTR engine emits usable structured text or one merged block.
 
-- **Homepage:** [https://doi.org/10.5281/zenodo.6827706](https://doi.org/10.5281/zenodo.6827706) 
-- **Paper:** [https://arxiv.org/abs/2207.11230](https://arxiv.org/abs/2207.11230)
+**The sibling dataset.** [`biglam/yalta_ai_segmonto_manuscript_dataset`](https://huggingface.co/datasets/biglam/yalta_ai_segmonto_manuscript_dataset) is the other half of the same experiment: manuscripts and early printed books, labelled with the [SegmOnto](https://segmonto.github.io/) zone vocabulary. This dataset does not use SegmOnto — tables were given their own four-class scheme.
 
-### Dataset Summary
+## Splits
 
-This dataset contains a subset of data used in the paper [You Actually Look Twice At it (YALTAi): using an object detectionapproach instead of region segmentation within the Kraken engine](https://arxiv.org/abs/2207.11230). This paper proposes treating page layout recognition on historical documents as an object detection task (compared to the usual pixel segmentation approach). This dataset covers pages with tabular information with the following objects "Header", "Col", "Marginal", "text". 
+|        | train | validation | test |
+|--------|-------|------------|------|
+| images | 196   | 22         | 135  |
 
-### Supported Tasks and Leaderboards
-
-- `object-detection`: This dataset can be used to train a model for object-detection on historic document images. 
-
-
-## Dataset Structure
-
-This dataset has two configurations. These configurations both cover the same data and annotations but provide these annotations in different forms to make it easier to integrate the data with existing processing pipelines. 
-
-- The first configuration, `YOLO`, uses the data's original format. 
-- The second configuration converts the YOLO format into a format which is closer to the `COCO` annotation format. This is done to make it easier to work with the `feature_extractor`s from the `Transformers` models for object detection, which expect data to be in a COCO style format. 
-
-### Data Instances
-
-An example instance from the COCO config:
-
-```
-{'height': 2944,
- 'image': <PIL.PngImagePlugin.PngImageFile image mode=L size=2064x2944 at 0x7FA413CDA210>,
- 'image_id': 0,
- 'objects': [{'area': 435956,
-   'bbox': [0.0, 244.0, 1493.0, 292.0],
-   'category_id': 0,
-   'id': 0,
-   'image_id': '0',
-   'iscrowd': False,
-   'segmentation': []},
-  {'area': 88234,
-   'bbox': [305.0, 127.0, 562.0, 157.0],
-   'category_id': 2,
-   'id': 0,
-   'image_id': '0',
-   'iscrowd': False,
-   'segmentation': []},
-  {'area': 5244,
-   'bbox': [1416.0, 196.0, 92.0, 57.0],
-   'category_id': 2,
-   'id': 0,
-   'image_id': '0',
-   'iscrowd': False,
-   'segmentation': []},
-  {'area': 5720,
-   'bbox': [1681.0, 182.0, 88.0, 65.0],
-   'category_id': 2,
-   'id': 0,
-   'image_id': '0',
-   'iscrowd': False,
-   'segmentation': []},
-  {'area': 374085,
-   'bbox': [0.0, 540.0, 163.0, 2295.0],
-   'category_id': 1,
-   'id': 0,
-   'image_id': '0',
-   'iscrowd': False,
-   'segmentation': []},
-  {'area': 577599,
-   'bbox': [104.0, 537.0, 253.0, 2283.0],
-   'category_id': 1,
-   'id': 0,
-   'image_id': '0',
-   'iscrowd': False,
-   'segmentation': []},
-  {'area': 598670,
-   'bbox': [304.0, 533.0, 262.0, 2285.0],
-   'category_id': 1,
-   'id': 0,
-   'image_id': '0',
-   'iscrowd': False,
-   'segmentation': []},
-  {'area': 56,
-   'bbox': [284.0, 539.0, 8.0, 7.0],
-   'category_id': 1,
-   'id': 0,
-   'image_id': '0',
-   'iscrowd': False,
-   'segmentation': []},
-  {'area': 1868412,
-   'bbox': [498.0, 513.0, 812.0, 2301.0],
-   'category_id': 1,
-   'id': 0,
-   'image_id': '0',
-   'iscrowd': False,
-   'segmentation': []},
-  {'area': 307800,
-   'bbox': [1250.0, 512.0, 135.0, 2280.0],
-   'category_id': 1,
-   'id': 0,
-   'image_id': '0',
-   'iscrowd': False,
-   'segmentation': []},
-  {'area': 494109,
-   'bbox': [1330.0, 503.0, 217.0, 2277.0],
-   'category_id': 1,
-   'id': 0,
-   'image_id': '0',
-   'iscrowd': False,
-   'segmentation': []},
-  {'area': 52,
-   'bbox': [1734.0, 1013.0, 4.0, 13.0],
-   'category_id': 1,
-   'id': 0,
-   'image_id': '0',
-   'iscrowd': False,
-   'segmentation': []},
-  {'area': 90666,
-   'bbox': [0.0, 1151.0, 54.0, 1679.0],
-   'category_id': 1,
-   'id': 0,
-   'image_id': '0',
-   'iscrowd': False,
-   'segmentation': []}],
- 'width': 2064}
-```
-
-An example instance from the YOLO config: 
-
-``` python
-{'image': <PIL.PngImagePlugin.PngImageFile image mode=L size=2064x2944 at 0x7FAA140F2450>,
- 'objects': {'bbox': [[747, 390, 1493, 292],
-   [586, 206, 562, 157],
-   [1463, 225, 92, 57],
-   [1725, 215, 88, 65],
-   [80, 1688, 163, 2295],
-   [231, 1678, 253, 2283],
-   [435, 1675, 262, 2285],
-   [288, 543, 8, 7],
-   [905, 1663, 812, 2301],
-   [1318, 1653, 135, 2280],
-   [1439, 1642, 217, 2277],
-   [1737, 1019, 4, 13],
-   [26, 1991, 54, 1679]],
-  'label': [0, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1]}}
-```
-
-
-
-### Data Fields
-
-The fields for the YOLO config:
-
-- `image`: the image
-- `objects`: the annotations which consist of:
-    - `bbox`: a list of bounding boxes for the image
-    - `label`: a list of labels for this image
-
-The fields for the COCO config:
-
-- `height`: height of the image
-- `width`: width of the image
-- `image`: image 
-- `image_id`: id for the image
-- `objects`: annotations in COCO format, consisting of a list containing dictionaries with the following keys:
-  - `bbox`: bounding boxes for the images
-  - `category_id`: a label for the image
-  - `image_id`: id for the image
-  - `iscrowd`: COCO `iscrowd` flag
-  - `segmentation`: COCO segmentation annotations (empty in this case but kept for compatibility with other processing scripts)
-
-
- 
-### Data Splits
-
-The dataset contains a train, validation and test split with the following numbers per split:
-
-
-|          | train | validation | test |
-|----------|-------|------------|------|
-| examples | 196   | 22         | 135  |
-
-
-## Dataset Creation
-
-> [this] dataset was produced using a single source, the Lectaurep Repertoires dataset [Rostaing et al., 2021], which served as a basis for only the training and development split. The testset is composed of original data, from various documents, from the 17th century up to the early 20th with a single soldier war report. The test set is voluntarily very different and out of domain with column borders that are not drawn nor printed in certain cases, layout in some kind of masonry layout. p.8
-.
-### Curation Rationale
-
-This dataset was created to produce a simplified version of the [Lectaurep Repertoires dataset](https://github.com/HTR-United/lectaurep-repertoires), which was found to contain:
-
-> around 16 different ways to describe columns, from Col1 to Col7, the case-different col1-col7 and finally ColPair and ColOdd, which we all reduced to Col p.8
-
-
-
-### Source Data
-
-#### Initial Data Collection and Normalization
-
-The LECTAUREP (LECTure Automatique de REPertoires) project, which began in 2018, is a joint initiative of the Minutier central des notaires de Paris, the National Archives and the 
- Minutier central des notaires de Paris of the National Archives, the [ALMAnaCH (Automatic Language Modeling and Analysis & Computational Humanities)](https://www.inria.fr/en/almanach) team at Inria and the EPHE (Ecole Pratique des Hautes Etudes), in partnership with the Ministry of Culture.
-
-> The lectaurep-bronod corpus brings together 100 pages from the repertoire of Maître Louis Bronod (1719-1765), notary in Paris from December 13, 1719 to July 23, 1765. The pages concerned were written during the years 1742 to 1745.
-
-#### Who are the source language producers?
-
-[More information needed]
-
-### Annotations
+Object instances per class, from Table 3 of the paper:
 
 |          | Train | Dev | Test | Total | Average area | Median area |
-|----------|-------|-----|------|-------|--------------|-------------|
-| Col      | 724   | 105 | 829  | 1658  | 9.32         | 6.33        |
-| Header   | 103   | 15  | 42   | 160   | 6.78         | 7.10        |
-| Marginal | 60    | 8   | 0    | 68    | 0.70         | 0.71        |
-| Text     | 13    | 5   | 0    | 18    | 0.01         | 0.00        |
-|          |       |     | -    |       |              |             |
+|----------|------:|----:|-----:|------:|-------------:|------------:|
+| Col      |   724 | 105 |  829 |  1658 |         9.32 |        6.33 |
+| Header   |   103 |  15 |   42 |   160 |         6.78 |        7.10 |
+| Marginal |    60 |   8 |    0 |    68 |         0.70 |        0.71 |
+| text     |    13 |   5 |    0 |    18 |         0.01 |        0.00 |
 
+> [!WARNING]
+> **Only `Col` and `Header` are annotated in the test split.** The paper's reason: "marginal and text were not clearly defined". `Marginal` and `text` have zero test instances and cannot be evaluated on held-out data — treat them as training signal only, and expect a model trained on all four classes to be scored on two.
 
-#### Annotation process
+> [!WARNING]
+> **The test split is out of domain by design.** Train and validation come from a single source; the test set is unrelated material from the 17th to the early 20th century, including one soldier's war report, with column borders that are neither drawn nor printed and layouts the paper calls "masonry". Test scores here measure transfer, not in-domain accuracy. In the paper the best model reaches 4.77% mAP on this test set, against 47.75% on the manuscripts dataset.
 
-[More information needed]
+Column labels were collapsed before release. The source annotations used "around 16 different ways to describe columns, from Col1 to Col7, the case-different col1-col7 and finally ColPair and ColOdd" — a convention that works around Kraken's tendency to merge adjacent regions of the same type. All of them were reduced to a single `Col`.
 
-#### Who are the annotators?
+## Source data
 
-[More information needed]
+Train and validation come from the [Lectaurep Repertoires](https://github.com/HTR-United/lectaurep-repertoires) dataset (Rostaing et al., 2021). LECTAUREP (LECTure Automatique de REPertoires), begun in 2018, is a joint project of the Minutier central des notaires de Paris at the Archives nationales, the [ALMAnaCH](https://www.inria.fr/en/almanach) team at Inria and the EPHE, in partnership with the French Ministry of Culture. The material is notaries' repertories: ruled, tabular, handwritten registers.
 
-### Personal and Sensitive Information
+## Loading the data
 
-This data does not contain information relating to living individuals.
-
-## Considerations for Using the Data
-
-### Social Impact of Dataset
-
-A growing number of datasets are related to page layout for historical documents. This dataset offers a different approach to annotating these datasets (focusing on object detection rather than pixel-level annotations). Improving document layout recognition can have a positive impact on downstream tasks, in particular Optical Character Recognition.
-
-### Discussion of Biases
-
-Historical documents contain a wide variety of page layouts. This means that the ability of models trained on this dataset to transfer to documents with very different layouts is not guaranteed. 
-
-### Other Known Limitations
-
-[More information needed]
+This repo still ships a legacy loading script, which `datasets` no longer runs:
 
+```python
+>>> load_dataset("biglam/yalta_ai_tabular_dataset", "YOLO")
+RuntimeError: Dataset scripts are no longer supported, but found yalta_ai_tabular_dataset.py
+```
 
-## Additional Information
+(checked against `datasets` 5.0.0). Until this repo is converted, take the data from Zenodo — a 359 MB zip laid out as `yaltai-table/{train,val,test}/{images,labels}`, one YOLO-format `.txt` per image (class index, then normalised centre-x, centre-y, width, height):
 
-### Dataset Curators
+```bash
+curl -L -o yaltai-table.zip "https://zenodo.org/record/6827706/files/yaltai-table.zip?download=1"
+```
 
+The loading script defined two configurations over the same annotations: `YOLO`, the original format, and `COCO`, converted for the Transformers object-detection processors. The COCO conversion left `segmentation` empty — these are boxes, not polygons.
 
-### Licensing Information
+## Licence
 
-[Creative Commons Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0/legalcode)
+[Creative Commons Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0/legalcode).
 
-### Citation Information
+## Citation
 
 ```
 @dataset{clerice_thibault_2022_6827706,
@@ -316,12 +90,8 @@ Historical documents contain a wide variety of page layouts. This means that the
 }
 ```
 
-
-
 [![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.6827706.svg)](https://doi.org/10.5281/zenodo.6827706)
 
+The accompanying paper: Thibault Clérice, "You Actually Look Twice At it (YALTAi): using an object detection approach instead of region segmentation within the Kraken engine", *Journal of Data Mining and Digital Humanities*, 2022. [arXiv:2207.11230](https://arxiv.org/abs/2207.11230).
 
-    
-### Contributions
-
-Thanks to [@davanstrien](https://github.com/davanstrien) for adding this dataset.
+Thanks to [@davanstrien](https://github.com/davanstrien) for adding this dataset to the Hub.