AvoCahDoe/region-proposal-pp-doclayout-l-bo

Morocco BO Region Proposal — PP-DocLayout-L Fine-tuned

Fine-tuned PP-DocLayout-L (RT-DETR-L, ~31M params / ~124 MB) on Moroccan Bulletin Officiel layout annotations for class-aware region proposal.

Run tag: 24_07_new_method (24 Jul 2026)
Demo Space: morocco-bo-region-proposal-demo

Training summary

Metric	Value
Base model	PP-DocLayout-L
Run tag	24_07_new_method
Dataset	BO export (BO_export_last), 30 docs / 1,625 pages
Train samples (after rare-class oversampling)	1,847
Val samples	174
Classes	Official 23-class PP-DocLayout taxonomy
Best val mAP	0.701 (epoch 38)
Early stopped	True (patience 15)
Model size	~123 MB inference weights

Evaluation (val + test, IoU 0.5, class-agnostic)

Metric	Fine-tuned	Pretrained baseline
Precision	0.8952	0.1206
Recall	0.7100	0.4085
F1	0.7919	0.1863

Usage (PaddleOCR)

from huggingface_hub import snapshot_download
from paddleocr import LayoutDetection

model_dir = snapshot_download(
    "AvoCahDoe/region-proposal-pp-doclayout-l-bo",
    allow_patterns=["inference/*"],
) + "/inference"

det = LayoutDetection(
    model_dir=model_dir,
    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
metadata/            # training config, logs, evaluation JSON
metrics/             # mirrored early_stop / metrics / eval JSON
plots/               # training + comparison plots
evaluation/plots/    # before/after overlays and training curves

Classes

23 PP-DocLayout labels including text, doc_title, paragraph_title, table, header, footer, reference, image, and more. See inference/inference.yml and metadata/config/pp_doclayout_categories.yaml.