Fine-tuned YOLOv8m for 8-class invoice document layout region detection (metadata, vendor/customer blocks, table, line items, summary, payment, column headers).

Part of the Hub collection Invoice Layout Extraction.

Property	Value
Architecture	YOLOv8m (25.9M params, 78.9 GFLOPs)
Base weights	`ultralytics/yolov8m`
Input size	1024 px
Classes	8 layout regions
Training epochs	100 (best @ epoch 98)
Optimizer	AdamW
Dataset	AvoCahDoe/invoice-annotated-bbox
Demo Space	AvoCahDoe/invoice-layout-yolov8n-demo
Train / val / test	372 / 14 / 7 pages

Classes

ID	Name
0	`invoice_metadata`
1	`vendor_block`
2	`customer_block`
3	`table_block`
4	`line_item`
5	`summary_block`
6	`payment_block`
7	`Column`

Metrics

Test split (held-out, 7 images)

Metric	Value
mAP50	0.8388
mAP50-95	0.5289
Precision	0.9077
Recall	0.8580

Validation (best epoch 98)

Metric	Best	Final (epoch 100)
mAP50	0.8581	0.8578
mAP50-95	0.5554	0.5544
Precision	0.8388	0.8253
Recall	0.7730	0.7651

Training time: ~66.6 min on RTX 4070 Laptop GPU.

Training configuration

Parameter	Value
Batch	6
Image size	1024
Patience	25
LR (cosine)	0.01 → 0.01
Mosaic	1.0
Mixup	0.15
Copy-paste	0.1
Horizontal flip	0.0

Training results

Learning curves

Precision–Recall

Confusion matrices

Raw	Normalized
Confusion matrix	Normalized confusion matrix

Validation predictions

Ground truth	Model predictions
Val labels	Val predictions

Training batches (augmented)

Batch 0	Batch 1	Batch 2
Train batch 0	Train batch 1	Train batch 2

Label distribution

Usage

Ultralytics (recommended)

python

from ultralytics import YOLO
from huggingface_hub import hf_hub_download

weights = hf_hub_download("AvoCahDoe/invoice-layout-yolov8m", "weights/best.pt")
model = YOLO(weights)
results = model.predict("invoice_page.png", imgsz=1024, conf=0.25)
results[0].show()

Load from Hub by repo id

python

from ultralytics import YOLO

model = YOLO("hf://AvoCahDoe/invoice-layout-yolov8m/weights/best.pt")
results = model.predict("invoice_page.png")

Training reproduction

python scripts/train_comparison.py --models yolov8m

Repository layout

weights/best.pt          # Best checkpoint (use this)
weights/last.pt          # Last epoch checkpoint
config/                  # Training configuration
metrics/                 # Per-epoch and summary metrics
assets/                  # Plots and visualizations

Model comparison (test split)

All six architectures fine-tuned on the same invoice layout dataset.

Rank	Model	test mAP50	test mAP50-95	Precision	Recall	Hub
1	`yolov8n`	0.9600	0.7167	0.9502	0.9592	YOLOv8n
2	`yolov8s`	0.9006	0.6739	0.9419	0.8677	YOLOv8s
3	`yolov8x`	0.8757	0.6224	0.9144	0.8735	YOLOv8x
4	`yolo11x`	0.8738	0.6127	0.9484	0.8530	YOLO11x
5	`yolo11m`	0.8418	0.4926	0.9033	0.8466	YOLO11m
6	`yolov8m` ← this model	0.8388	0.5289	0.9077	0.8580	YOLOv8m

License

Apache 2.0