AvoCahDoe/sommaire-pp-doclayout-l-proposal

SommaireTOC Region Proposal — PP-DocLayout-L (gated)

Fine-tuned PaddlePaddle/PP-DocLayout-L for document region proposal on French Bulletin Officiel (BO) pages, used in the SommaireTOC pipeline (SOMMAIRE / TOC extraction).

Gated model — trained on proprietary/internal BO annotations. Inference export from run sommaire_proposal_new_dataset_08-08-2026.

Model description

Property	Value
Architecture	PP-DocLayout-L (DETR), Paddle inference
Base model	PaddlePaddle/PP-DocLayout-L
Task	Multi-class document layout / region proposal
Metric (selection)	COCO bbox mAP
Best checkpoint	Epoch 16, mAP 0.7039
Input size	640 × 640
Framework	PaddlePaddle / PaddleOCR LayoutDetection

Label taxonomy (23 PP-DocLayout classes)

paragraph_title, image, text, number, abstract, content, figure_title, formula, table, table_title, reference, doc_title, footnote, header, algorithm, footer, seal, chart_title, chart, formula_number, header_image, footer_image, aside_text

Training data

Item	Value
Pages	775
Documents	155
Domain	French Bulletin Officiel (SOMMAIRE / front-matter heavy)
Run id	sommaire_proposal_new_dataset_08-08-2026

Training / checkpoint selection

Item	Value
Best val COCO mAP	0.703857
Best epoch	16
Export	best_model/inference (inference.pdiparams + inference.json + inference.yml)

Full training ledger: metrics/train_result.json
Best info: metrics/best_model.info.json

Evaluation results

Validation (training selection)

Metric	Value
Best COCO mAP	0.7039
Best epoch	16

Held-out page eval (n=120) vs pretrained baseline

From metrics/eval_metrics.json:

Metric	Finetuned	Baseline
Macro F1	0.9865	(see JSON)
Weighted F1	0.9874	(see JSON)
Mean matched IoU	0.8378	(see JSON)

Per-class F1 (finetuned, matched evaluation)

Class	Precision	Recall	F1	Support
header	0.984	0.974	0.979	190
paragraph_title	0.991	0.997	0.994	921
reference	0.990	1.000	0.995	96
text	0.961	0.997	0.978	593

Plots

Sample detections

Example pages

Usage (PaddleOCR)

from huggingface_hub import snapshot_download
from paddleocr import LayoutDetection

model_dir = snapshot_download(
    "AvoCahDoe/sommaire-pp-doclayout-l-proposal",
    allow_patterns=["inference/*"],
)
# Hub layout keeps weights under inference/
det = LayoutDetection(
    model_dir=f"{model_dir}/inference",
    model_name="PP-DocLayout-L",
    enable_mkldnn=False,
)
output = det.predict("page.png", layout_nms=True, threshold=0.5)
for res in output:
    data = res.json if hasattr(res, "json") else {}
    for box in data.get("res", data).get("boxes", []):
        print(box.get("label"), box.get("score"), box.get("coordinate"))

Repository layout

inference/   # Exported Paddle inference model (use this at runtime)
metrics/     # train_result, best_model.info, eval_metrics
plots/       # training / validation curves
samples/     # detection overlay visualizations
examples/    # raw BO page examples

Intended use

• Region proposal stage of SommaireTOC on French BO PDFs
• Upstream of LayoutLMv3 segment classification (AvoCahDoe/sommaire-layoutlmv3-classifier)

Limitations

• Tuned for Moroccan/French Bulletin Officiel layouts; other document types may need further finetuning
• Private weights; do not redistribute without authorization
• Stock PP-OCR det/rec are not included (use separately)

Citation

If you use this model in work derived from the SommaireTOC pipeline, please cite the Hub repo and base model PaddlePaddle/PP-DocLayout-L.