# LADaS 2.0: Diachronic Document Dataset for Semantic Layout Analysis

> LADaS 2.0 is a diachronic, multi-domain document layout analysis dataset of 7,254 manually annotated pages spanning 1600 to 2024, using a SegmOnto/TEI-aligned taxonomy of 36 classes across 13 zone types.

- Authors: Bevaya AI Team
- Published: 2026-03-10
- Updated: 2026-03-12
- Paper: Preprint (arXiv) <https://arxiv.org/abs/2411.10068>
- Dataset: LADaS 2.0 Dataset (Hugging Face) <https://huggingface.co/datasets/almanach/LADaS>
- Code: LADaS Code & Tools <https://github.com/DEFI-COLaF/LADaS>
- Source: Bevaya Labs, <https://labs.bevaya.ai/>

---


## TL;DR

LADaS 2.0 is an open, diachronic document layout analysis dataset containing 7,254 manually annotated pages from the 17th century to the present. It uses a 36-class taxonomy (13 zone types with subtypes) grounded in the SegmOnto vocabulary and mapped to the Text Encoding Initiative (TEI) standard, making it one of the few DLA datasets designed with document reconstruction workflows in mind. The authors benchmark YOLOv11 across multiple input sizes and model scales, finding that 1280-pixel input with a large model performs best, and that training on the full generic dataset generally outperforms subset-specific fine-tuning.

## What kind of paper is this?

**Dominant:** $\Psi\_{\text{Resource}}$ : the paper's headline contribution is the dataset itself: its annotation guidelines, class taxonomy, subset composition, and release as an open resource. The curation pipeline, TEI/SegmOnto alignment, and metadata design occupy the majority of the paper.

**Secondary:** $\Psi\_{\text{Evaluation}}$ : the paper benchmarks YOLOv11 models at three input resolutions and five model scales, plus a comparison against DocYOLO. A second experiment evaluates domain-specific fine-tuning vs. generic training on held-out subsets.

## What is the motivation?

Document layout analysis datasets have historically skewed toward born-digital, English-language STEM papers (PubLayNet, DocBank) or narrow modern document types (DocLayNet). Cultural heritage institutions and digital humanities (DH) researchers work with materials that span centuries of typographic evolution, yet few DLA datasets capture this temporal and structural diversity.

Three specific gaps motivate LADaS 2.0:

1. **Diachronic coverage.** Existing datasets either target a single era or a single document type. Historical documents exhibit layout conventions (marginal notes, running titles, ornamental graphics, verse structures) that differ substantially from modern PDF layouts. A dataset spanning 1600 to 2024 forces models to generalize across these conventions.

2. **Semantic granularity aligned to document reconstruction.** Most DLA datasets use a flat label set (text, title, table, figure). DH workflows need finer distinctions: paragraphs vs. verse groups vs. list items vs. dramatic speeches within the main text zone, as well as separation of footnotes from manuscript additions in margin zones. The SegmOnto vocabulary provides a principled first level, but it lacks subtype specifications needed for full TEI reconstruction.

3. **Interoperability.** The DH community adopted SegmOnto as a shared vocabulary, enabling datasets like Gallicorpora and Ajax to be combined. LADaS 2.0 extends this by defining subtypes that map directly to TEI XML elements, bridging the gap between CV-style bounding box annotations and structured text encoding.

## What is the novelty?

The core contribution is the dataset and its taxonomy. Several design choices distinguish it from prior work:

**TEI-grounded subtype system.** LADaS 2.0 extends SegmOnto's zone types with subtypes selected based on three criteria: (1) a corresponding TEI XML element exists, (2) the element is visually distinguishable in the page image, and (3) the distinction is relevant for downstream document reconstruction. This yields 36 classes organized under 13 zone types. For example, `MainZone` is subdivided into `Head`, `P` (paragraph), `Lg` (line group/verse), `Sp` (dramatic speech), `List`, `Entry`, `Date`, `Signature`, `Maths`, and `Other`.

**Modular subset architecture.** The 7,254 pages are organized into 12 subsets by provenance and content type (Monographies, Catalogues, Theses, Theatre, Magazines, etc.), each with metadata including publication date, domain, and acquisition method. This modularity enables users to compose domain-specific training sets or study cross-domain transfer.

**Diachronic metadata.** 95% of documents carry a publication year, enabling temporal analysis. The dataset spans from the 17th century (primarily theatre texts from Gallica) to born-digital 21st-century theses and administrative reports.

**"Continued" subtype for run-on elements.** When a paragraph or entry spans a page break, the continuation on the next page is annotated as `[Zone]-Continued`. During reconstruction, reading-order heuristics merge these with their predecessor, maintaining document coherence.

**New SegmOnto zone types.** The authors introduce `FigureZone` (for code excerpts), `FormZone` (for magazine forms), and `AdvertisementZone` (for advertisements), types that were absent from the original SegmOnto vocabulary.

**Noise subset (Fingers).** A deliberately noisy subset of 100 pages was created by scanning books with visible fingers, background clutter, and bent pages, simulating real-world on-site digitization conditions.

## What experiments were performed?

### Generic model benchmarks

The authors train YOLOv11 models at three input resolutions (640, 960, 1280 pixels) across five model scales (nano, small, medium, large, extra-large), for 15 configurations total. Three subsets (Theatre, Admin. Rep., Romans-19) are held out from training for the domain-specific experiments. Training uses the Ultralytics library (v8.3.8) with batch size 16, 100 epochs, learning rate 0.01, seed 42, and standard augmentations (rotation, contrast, shear). A single RTX 8000 GPU is used for all models except the 1280-pixel extra-large variant, which requires two.

DocYOLO (a layout-specific adaptation of YOLOv10) is also evaluated as a comparison, trained on two GPUs with the same parameters except a learning rate of 0.02.

The primary metric is mAP@50 (COCO-style).

### Domain-specific fine-tuning

For each of the three held-out subsets, the authors train three models: (1) generic training with the subset added to the full pool, (2) fine-tuning from raw YOLOv11-L weights on the subset alone, and (3) fine-tuning from the generic model (Exp-1) on the subset alone. The baseline is the generic model that never saw the subset.

## What are the outcomes/conclusions?

**Input resolution matters.** The 1280-pixel input consistently outperforms 960, which outperforms 640. This aligns with annotator experience: marginal notes and small zones are difficult to distinguish at low resolution.

**Model scale plateaus.** Performance gains diminish beyond the large model. The extra-large model provides minimal improvement and occasionally degrades, possibly due to overfitting given the dataset size. The **YOLOv11-L at 1280 pixels** achieves the best overall mAP@50.

**DocYOLO underperforms.** Despite layout-specific architectural modifications, DocYOLO scores below the standard YOLOv11-L, suggesting that input resolution and dataset diversity may matter more than architectural tweaks for this task.

**Generic training beats fine-tuning.** For 2 of 3 held-out subsets (Theatre, Romans-19), the generic model incorporating the subset data outperforms fine-tuning on the subset alone. The Admin. Rep. subset is a marginal exception (fine-tuned Exp-1-L edges out the generic model by less than one point on mAP@50). Fine-tuning from raw YOLOv11-L weights on small subsets consistently performs worst.

**Challenging subsets.** The Fingers (noisy scans) and Tech. Magazines (high visual complexity, limited training data) subsets yield the lowest per-subset scores, suggesting that data augmentation strategies targeting noise and complex layouts could help.

The dataset contains 81,766 total annotation instances across 7,254 pages. The most common classes (`MainZone-P`, `MainZone-Head`) are well-represented across the full temporal range, while niche classes (`QuireMarks`, `NumberingZone`) are sparser but consistently present from the 18th century onward.

## Mapping to Matter vs. Meaning Framework

How does LADaS 2.0's SegmOnto/TEI-aligned taxonomy map to our [Matter vs. Meaning framework](../matter-vs-meaning/)? LADaS uses a two-level hierarchy: zone types (from SegmOnto) with subtypes (mapped to TEI elements).

### MainZone (Body Content)

| LADaS Class | Visual Primitive | Logical Role | Notes |
| :--- | :--- | :--- | :--- |
| **MainZone-Head** | `Text` | `SectionHeader` | Section headings within the main text. |
| **MainZone-P** | `Text` | `Body` | Standard paragraphs. The most common class. |
| **MainZone-Lg** | `Text` | `Body` | Line groups (verse, poetry). Preserves line breaks as semantic structure. |
| **MainZone-Sp** | `Text` | `Body` | Dramatic speeches (theatre texts). Domain-specific. |
| **MainZone-List** | `Text` | `ListItem` | List structures. |
| **MainZone-Entry** | `Text` | `Body` | Dictionary/catalogue entries. |
| **MainZone-Date** | `Text` | `Dateline` | Date elements in the main flow. |
| **MainZone-Signature** | `Text` | `Signature` | Author signatures at end of sections. |
| **MainZone-Maths** | `Formula` | `DisplayEquation` | Mathematical content in the main zone. |
| **MainZone-Other** | `Text` | `Body` | Catch-all for main content not fitting other subtypes. |

### MarginTextZone

| LADaS Class | Visual Primitive | Logical Role | Notes |
| :--- | :--- | :--- | :--- |
| **MarginTextZone-Notes** | `Text` | `Footnote` | Marginal notes (printed). |
| **MarginTextZone-ManuscriptAddendum** | `Text` | `Annotation` | Handwritten additions in margins. `is_handwritten: True`. |

### Visual & Graphic Zones

| LADaS Class | Visual Primitive | Logical Role | Notes |
| :--- | :--- | :--- | :--- |
| **FigureZone** | `Text` | `Code` | Code excerpts (repurposed zone type). |
| **FigureZone-Head** | `Text` | `Caption` | Heading/caption for code blocks. |
| **FigureZone-Figdesc** | `Text` | `Caption` | Description of code figures. |
| **GraphicZone** | `Image` | `Figure` | Visual content (illustrations, photographs). |
| **GraphicZone-Head** | `Text` | `Caption` | Heading for graphic zones. |
| **GraphicZone-Figdesc** | `Text` | `Caption` | Description of graphics. |
| **GraphicZone-TextualContent** | `Text` | `Body` | Text embedded within graphic regions. |
| **GraphicZone-Part** | `Image` | `Figure` | Sub-components of a larger graphic. |
| **GraphicZone-Decoration** | `Structure` | *(decorative)* | Ornamental graphics (vignettes, borders). |

### Structural & Navigation Zones

| LADaS Class | Visual Primitive | Logical Role | Notes |
| :--- | :--- | :--- | :--- |
| **TableZone** | `Table` | *(primitive only)* | Tabular structures. |
| **TableZone-Head** | `Text` | `Caption` | Table headings/captions. |
| **PageTitleZone** | `Text` | `Title` | Page-level titles. |
| **PageTitleZone-Index** | `Text` | `Title` | Index page titles. |
| **RunningTitleZone** | `Text` | `PageHeader` | Running titles in margins/headers. |
| **NumberingZone** | `Text` | `PageNumber` | Page and element numbering. |
| **StampZone** | `Image` | `Stamp` | Library stamps, ownership marks. |
| **StampZone-Sticker** | `Image` | `Stamp` | Adhesive labels. |
| **FormZone** | `Table` | `FormGroup` | Form structures (magazines). |
| **AdvertisementZone** | `Image` | `Advertisement` | Commercial advertisements. |

### Continued Variants

LADaS introduces a `-Continued` suffix for elements spanning page breaks (e.g., `MainZone-P-Continued`, `MainZone-Entry-Continued`). During reconstruction, reading-order heuristics merge these with their predecessor on the previous page. This addresses a common gap in single-page datasets that cannot represent cross-page continuity.

**Coverage strengths:** LADaS is one of the few datasets designed with **document reconstruction** (TEI output) as the explicit goal. The SegmOnto/TEI alignment means every class maps to a specific XML element. The diachronic coverage (1600-2024) and domain-specific labels (dramatic speeches, catalogue entries, manuscript addenda) make it uniquely suited for cultural heritage workflows. The `FormZone` and `AdvertisementZone` are new additions to SegmOnto.

**Coverage gaps:** No explicit classes for `Author`, `Affiliation`, `Abstract`, `BibEntry`, `Footnote` (separate from margin notes), `ListItem` (individual items), or any form interaction primitives (`Field`, `Selection`, `Checkbox`). The dataset is heavily French-language.

## Reproducibility

### Models

No new architecture is proposed. All experiments use standard YOLOv11 (Ultralytics) and DocYOLO:
- YOLOv11: nano (n), small (s), medium (m), large (l), extra-large (x) variants.
- DocYOLO: layout-adapted YOLOv10.
- No trained weights are released with the paper.

### Algorithms

- **Training:** Ultralytics library v8.3.8. Batch size 16, 100 epochs, learning rate 0.01 (0.02 for DocYOLO), seed 42.
- **Augmentations:** Rotations, contrast, shear (Ultralytics defaults).
- **Evaluation metric:** mAP@50 (COCO-style).
- All other parameters are Ultralytics defaults.

### Data

- 7,254 annotated pages; train: 5,443, validation: 912, test: 899.
- 12 subsets organized by provenance and content type.
- Annotations in YOLOv8 txt format (`class center_x center_y width height`, normalized coordinates).
- The HuggingFace release contains 6,060 rows (some images may be excluded from the HF version relative to the paper's count; the GitHub repository is the primary source for the full dataset).
- 36 classes across 13 zone types; full taxonomy in Table 2 of the paper.
- License: CC-BY-4.0 for annotations and code.
- Underlying document licenses vary by subset: Gallica materials are public domain, ARLP (Picard) and INHA/BnF (Catalogues) are partner donations, Persee materials are from the Persee portal, and theses are from theses.fr. The CC-BY-4.0 applies to the annotations; users should verify source-document terms for their use case.
- Annotation process: manual annotation on Roboflow, with pre-annotations from iteratively trained models. Each annotation reviewed by a second annotator; disagreements resolved by guideline authors.

### Evaluation

- mAP@50 reported per model configuration and per subset.
- No error bars, confidence intervals, or multi-seed runs reported for the main experiments. The authors note that additional seed tests were conducted for the extra-large model plateau phenomenon.
- Three held-out subsets (Theatre, Admin. Rep., Romans-19) enable controlled domain-transfer evaluation.
- Comparison against DocYOLO uses the same hyperparameters except learning rate.

### Hardware

- Single NVIDIA RTX 8000 GPU for most models.
- Two RTX 8000 GPUs for the 1280-pixel extra-large model and for DocYOLO.
- No wall-clock training times, VRAM usage, or cost estimates reported.

## BibTeX

```bibtex
@misc{clerice2024diachronic,
  title={Diachronic Document Dataset for Semantic Layout Analysis},
  author={Thibault Clérice and Juliette Janès and Hugo Scheithauer and Sarah Bénière and Florian Cafiero and Laurent Romary and Simon Gabay and Benoît Sagot},
  year={2024},
  eprint={2411.10068},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}
```