AvoCahDoe/layoutlmv3-bo-segments

Demo Space: morocco-bo-region-classification-demo

LayoutLMv3 — Bulletin Officiel Segment Classifier

Fine-tuned microsoft/layoutlmv3-base for segment-level layout classification on French Bulletin Officiel (BO) documents.

Each annotated region (bounding box + OCR words inside) is encoded as a single sequence; the model predicts one of 14 canonical layout classes for the whole region (LayoutLMv3ForSequenceClassification).

Private model — trained on proprietary/internal BO annotations. See Training data below.

Model description

Property	Value
Architecture	LayoutLMv3ForSequenceClassification
Base model	microsoft/layoutlmv3-base
Task	Segment-level sequence classification
Classes	14 canonical BO layout roles
Input	Page image + word tokens + word bounding boxes (normalized 0–1000)
Max sequence length	512
Best checkpoint selection	f1_macro on validation set
Training run	run_segments_20260619_015948 (early stopping, 19/60 epochs)

Layout classes (14)

ARTICLE, PREAMBLE, TITLE, FOOTER, TABLE, FIGURE, ANNEXE_TITLE, ANNEXE_TEXT, SOMMAIRE, CHAPTER_CONTENT, ANNEXE_LEVEL, CHAPTER_TITLE, FORM, SECTION

Training data

Fine-tuned on a French Bulletin Officiel document layout dataset built in the layoutlmv3-bo pipeline.

Source documents (document-level split, no page leakage)

Document	Pages	Split
BO_7492_fr.pdf	90	train
BO_7506_Fr.pdf	100	train
BO_7496_Fr.pdf	13	train
BO_7514_Fr.pdf	4	train
BO_7510_Fr.pdf	32	val
BO_7500_Fr.pdf	26	test

Segment statistics

Split	Segments
Train	783 (+ augmentation)
Val	221
Test	147

• 265 annotated pages across 6 PDFs
• 1,172 annotated layout regions (bounding boxes)
• 126,579 OCR words (PyMuPDF + EasyOCR hybrid on sparse pages)
• Words inside each GT region form one training example (≥2 words per segment)
• Augmentation: rare-class oversampling (×3), bbox jitter (±2 px)

Full dataset documentation: metadata/DATASET.md
Split manifest: metadata/split_manifest.json
Label schema: metadata/segment_label_schema.json

Training hyperparameters

Parameter	Value
Epochs requested	60
Epochs completed	19 (early stopping, patience 8)
Batch size (per device)	4
Gradient accumulation	2
Learning rate	2e-5
Weight decay	0.01
Warmup ratio	0.1
Optimizer metric	f1_macro
FP16	true
Class weights	inverse-frequency (segment_class_weights.json)

Training config: metadata/logs/run_config_segments_overnight.json
Per-epoch val metrics: metadata/logs/segments_overnight_20260619_015948_epochs.json
Step-level log: metadata/logs/training_log_segments.json

Evaluation results

Held-out test set (BO_7500_Fr.pdf, 147 segments)

Metric	Value
Accuracy	0.9728
F1 macro	0.8500
F1 micro	0.9728
Precision macro	0.8492
Recall macro	0.8535

Per-class F1 (test)

Class	F1
ARTICLE	1.000
FOOTER	1.000
SOMMAIRE	1.000
TABLE	1.000
TITLE	0.950
PREAMBLE	0.941
ANNEXE_TITLE	0.909
CHAPTER_TITLE	0.000 (0 support in test)

Full metrics: metadata/evaluation/metrics.json
Classification report: metadata/evaluation/classification_report.txt
Aligned predictions: metadata/evaluation/aligned_predictions.json

Evaluation plots

All plots live under evaluation/plots/ (22 files: 7 summary charts + 15 test-set page overlays on BO_7500_Fr.pdf).

Summary charts

Plot	File
Color legend (14 BO classes)	00_color_legend.png
Segments per class (dataset)	01_segments_per_class.png
Training / validation curves	01_training_curves.png
Confusion matrix (raw counts)	03_confusion_raw.png
Confusion matrix (normalized)	03_confusion_norm.png
Per-class F1 (test)	04_per_class_f1.png
Train segment distribution	05_train_segment_distribution.png

Test-set page overlays (BO_7500_Fr.pdf)

Predicted segment labels overlaid on held-out test pages (one color per BO class; see legend above).

Page	File
page 002	overlay_BO_7500_Fr__doc7__page_002.png
page 006	overlay_BO_7500_Fr__doc7__page_006.png
page 007	overlay_BO_7500_Fr__doc7__page_007.png
page 016	overlay_BO_7500_Fr__doc7__page_016.png
page 017	overlay_BO_7500_Fr__doc7__page_017.png
page 018	overlay_BO_7500_Fr__doc7__page_018.png
page 019	overlay_BO_7500_Fr__doc7__page_019.png
page 021	overlay_BO_7500_Fr__doc7__page_021.png
page 022	overlay_BO_7500_Fr__doc7__page_022.png
page 023	overlay_BO_7500_Fr__doc7__page_023.png
page 025	overlay_BO_7500_Fr__doc7__page_025.png
page 026	overlay_BO_7500_Fr__doc7__page_026.png
page 029	overlay_BO_7500_Fr__doc7__page_029.png
page 030	overlay_BO_7500_Fr__doc7__page_030.png
page 032	overlay_BO_7500_Fr__doc7__page_032.png

Usage

from transformers import LayoutLMv3ForSequenceClassification, LayoutLMv3Processor
from PIL import Image

model_id = "AvoCahDoe/layoutlmv3-bo-segments"
processor = LayoutLMv3Processor.from_pretrained("microsoft/layoutlmv3-base", apply_ocr=False)
model = LayoutLMv3ForSequenceClassification.from_pretrained(model_id)

# words: list[str], boxes: list[[x0,y0,x1,y1]] normalized 0-1000
image = Image.open("page.png").convert("RGB")
encoding = processor(
    image, words, boxes=boxes,
    return_tensors="pt", truncation=True, padding="max_length", max_length=512,
)
outputs = model(**{k: v for k, v in encoding.items()})
pred_id = outputs.logits.argmax(-1).item()
label = model.config.id2label[str(pred_id)]

Pipeline integration

Used downstream in the RegionProposal pipeline: PP-DocLayout proposes region boxes → this model classifies each region into BO taxonomy.

python scripts/run_layoutlmv3_on_proposals.py --run-dir runs/BO_7458_fr

Repository layout

config.json                 # Model config + id2label / label2id
model.safetensors           # Fine-tuned weights (~481 MB)
training_args.bin           # HuggingFace TrainingArguments snapshot
metadata/
  DATASET.md                # Full dataset documentation
  segment_label_schema.json
  split_manifest.json
  evaluation/               # Test metrics, reports, predictions
  logs/                     # Training config and curves
evaluation/plots/           # Confusion matrices, training curves, overlays

Limitations

• Requires pre-segmented regions (GT boxes or proposal boxes from PP-DocLayout); not an end-to-end detector.
• Trained on 6 French government PDFs — may not generalize to other layouts or languages.
• Rare classes (CHAPTER_TITLE, FORM, SECTION) have limited test support.
• OCR quality affects performance (PyMuPDF primary, EasyOCR fallback).

Citation

@misc{layoutlmv3-bo-segments-2026,
  title={LayoutLMv3 Fine-tuned for French Bulletin Officiel Segment Classification},
  author={AvoCahDoe},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/AvoCahDoe/layoutlmv3-bo-segments}}
}

Acknowledgments

• Base model: microsoft/layoutlmv3-base (LayoutLMv3, Microsoft)
• Training framework: Hugging Face Transformers