🔥 Latest Updates

• 2026/03/10: Released YOLOv26 models - Get YOLOv26x
• 2025/03/10: Released YOLOv12 models - Get YOLOv12x
• 2024/10/07: Released YOLOv11 models - Get YOLOv11x
• 2024/07/10: Released YOLOv10 models - Get YOLOv10x
• 2024/06/21: Released YOLOv9 models

🎯 Model Demo

📊 Performance Results

Model Performance Comparison Chart (mAP50-95)

Detailed Model Performance Metrics (Parameters/mAP50-95)

Refer to Detail Results

Why this repo?

RAG (Retrieval Augmented Generation) is widely used today for chatting with documents. But when documents have complex layouts, the performance often suffers. It's hard to properly extract and structure content from these complex documents. This project offers a fast and effective solution to this problem.

1. YOLO is a leading object detection model by Ultralytics. It comes in 5 different sizes and has a robust framework for training and deployment. I picked YOLO because of these strengths.

2. DocLayNet is a dataset of 80,863 document pages with human-labeled layout information. It includes many different types of documents and is currently the best dataset available for document layout analysis.

What I did?

Here's what I did:

1. Created a script that converts DocLayNet data into YOLO's training format
2. Built code for training, testing and running the models
3. Trained and shared YOLO models in all sizes and versions

How to use?

from ultralytics import YOLO

model = YOLO("{path to model file}")
pred = model("{path to test image}")
print(pred)

The definition of predict result please refer to the doc.

Server

You can simply python main.py to serve the model. Open http://localhost:8000/redoc check the API.

Dataset

DocLayNet can be found more details and download at this link. It has 11 labels:

• Text: Regular paragraphs.
• Picture: A graphic or photograph.
• Caption: Special text outside a picture or table that introduces this picture or table.
• Section-header: Any kind of heading in the text, except overall document title.
• Footnote: Typically small text at the bottom of a page, with a number or symbol that is referred to in the text above.
• Formula: Mathematical equation on its own line.
• Table: Material arranged in a grid alignment with rows and columns, often with separator lines.
• List-item: One element of a list, in a hanging shape, i.e., from the second line onwards the paragraph is indented more than the first line.
• Page-header: Repeating elements like page number at the top, outside of the normal text flow.
• Page-footer: Repeating elements like page number at the bottom, outside of the normal text flow.
• Title: Overall title of a document, (almost) exclusively on the first page and typically appearing in large font.

Prepare data

You can download the processed dataset directly from Hugging Face: hantian/doclaynet-for-yolo.

Alternatively, you can prepare the data manually:

1. download DocLayNet dataset by this link
2. unzip to datasets folder
3. use my convert script to make datasets ready for training

wget https://codait-cos-dax.s3.us.cloud-object-storage.appdomain.cloud/dax-doclaynet/1.0.0/DocLayNet_core.zip
mkdir datasets
mv DocLayNet_core.zip datasets/
cd datasets/ && unzip DocLayNet_core.zip && rm DocLayNet_core.zip
cd ../
python convert_dataset.py

Train & Eval

train

After preparing data, thanks to Ultralytics, training is super easy. You can choose base models from this link. I use the YOLOv8 series.

python train.py {base-model}

Eval

After training, you can evaluate your best model on test split.

python eval.py {path-to-your-model}

Detail Results

YOLOv26 Models

YOLOv12 Models

YOLOv11 Models

YOLOv10 Models

YOLOv9 Models

YOLOv8 Models