Donut Rus

🚀 Donut Model for Russian Text Recognition

Donut, an end-to-end transformer model, designed for recognizing Russian text.

🚀 Quick Start

The Donut model is an end-to-end transformer model for recognizing Russian text. You can access the official Donut repository here.

✨ Features

This model serves as a base for the following tasks:

• Recognizing various types of documents.
• Answering questions about documents.
• Classifying documents.

To solve your specific tasks, you can use the notebooks mentioned above. The dataset needs to be annotated in the format specified in the Donut repository.

📦 Installation

No installation steps were provided in the original document, so this section is skipped.

💻 Usage Examples

No code examples were provided in the original document, so this section is skipped.

📚 Documentation

Model Overview

The Donut model is an end-to-end transformer model for Russian text recognition. A custom dataset named SynthDoG, consisting of 100,000 images with texts taken from Russian literature works, was generated for training. You can access the dataset here.

The model was trained on a notebook by NielsRogge on Kaggle, with the original tokenizer replaced by DeepPavlov/xlm - roberta - large - en - ru. You can find the training notebook here.

The validation metric, Normed ED, is 0.02239.

Applicable Datasets

• Akajackson/donut_synthdog_rus

Supported Languages

• ru
• en

🔧 Technical Details

The model is an end - to - end transformer model. A custom dataset SynthDoG was generated for training, containing 100,000 images with texts from Russian literature. The training was conducted on a Kaggle notebook, replacing the original tokenizer with DeepPavlov/xlm - roberta - large - en - ru. The validation metric Normed ED is 0.02239, indicating good performance on the validation set.

📄 License

No license information was provided in the original document, so this section is skipped.