# Dataset Specification

## Overview

Dataset name: Detecting Latin in Historical Books with Large Language Models: A Multimodal Benchmark

Target Hugging Face repository: `COMHIS/eacl26-detect-latin`

Source record: https://doi.org/10.5281/zenodo.18377924

Source code: https://github.com/COMHIS/EACL26-detect-latin

Paper: https://aclanthology.org/2026.eacl-long.245/

Paper DOI: https://doi.org/10.18653/v1/2026.eacl-long.245

License: CC BY-NC 4.0

The dataset contains 724 annotated historical document pages. 594 pages contain Latin and 130 pages do not. The data is designed for page-level Latin detection and Latin segment extraction from OCR text and page images.

## Source Files

The Zenodo record contains:

| File | Size | Checksum |
| --- | ---: | --- |
| `latin_annotation_final.json` | 6,176,204 bytes | `md5:f7dd7b59c2ba90832bfa7f5e2fc16976` |
| `images.zip` | 52,835,690 bytes | `md5:a1f4d0e1ed78561e137ae276d75a89ff` |

## Recommended Hugging Face Layout

```text
README.md
DATASET_SPEC.md
data/
  pages.jsonl
  latin_annotation_final.json
images.zip
```

`pages.jsonl` is the normalized, viewer-friendly file. `latin_annotation_final.json` preserves the original annotation export. `images.zip` preserves the original image bundle.

## Normalized Record Schema

Each line in `data/pages.jsonl` is a JSON object:

| Field | Type | Description |
| --- | --- | --- |
| `page_id` | string | Stable page id, `<ecco_id>_<page_number>`. |
| `ecco_id` | string | ECCO source identifier. |
| `page_number` | integer or string | Page number from the source item. |
| `image_path` | string | Relative image path, e.g. `images/0213801900_151.jpg`. |
| `has_latin` | boolean | `true` when `page_latin` is non-empty. |
| `page_text` | string | Original OCR text from the `pageText` field. |
| `page_text_clean` | string | LLM-corrected OCR text from the `pageTextClean` field. |
| `page_latin` | string | Latin text associated with the page. |
| `latin_regions` | list[object] | Bounding boxes for Latin regions. |
| `latin_spans` | list[object] | Text-span labels for Latin segments. |

### `latin_regions`

Each region object has:

| Field | Type | Description |
| --- | --- | --- |
| `x` | float | Left coordinate as percentage of image width. |
| `y` | float | Top coordinate as percentage of image height. |
| `width` | float | Box width as percentage of image width. |
| `height` | float | Box height as percentage of image height. |
| `rotation` | float | Rotation value from the annotation tool. |
| `labels` | list[string] | Region labels. All released region annotations use `["Latin"]`. |
| `original_width` | integer | Original image width in pixels. |
| `original_height` | integer | Original image height in pixels. |

### `latin_spans`

Each span object has:

| Field | Type | Description |
| --- | --- | --- |
| `start` | integer | Start character offset in the annotated text field. |
| `end` | integer | End character offset in the annotated text field. |
| `text` | string | Annotated Latin text span. |
| `labels` | list[string] | Fine-grained Latin use labels. |

## Observed Label Distribution

Text-span labels:

| Label | Count |
| --- | ---: |
| Direct Quote | 389 |
| Independent | 230 |
| Footnote | 192 |
| Code-switching | 183 |
| Side-note | 81 |
| Bilingual | 55 |
| Emblematic | 38 |
| Legal | 31 |
| Indices and Catalogs | 30 |
| Ecclesiastical | 23 |
| Tables and Charts | 10 |
| Dictionary | 9 |
| Unlocatable | 5 |

The original export contains 2,310 Latin bounding-box annotations.

## Splits

No train/validation/test split is included in the released dataset. The release contains all 724 annotated pages as a single benchmark set. Any derived split should be reported by downstream users.

## Known Caveats

- OCR text is noisy, reflecting the historical print source and OCR pipeline.
- Image paths in the JSON are relative to the extracted `images/` directory.
- The data is licensed for non-commercial use only.