RuDialect-HTR is a project focused on the detection and recognition of handwritten Russian dialectic texts using modern deep learning models.
This repository contains four Jupyter notebooks:
fine-tuning-craft— fine-tuning the CRAFT model for text detectionfine-tuning-yolo— fine-tuning YOLO for text detectionfine-tuning-trocr— fine-tuning TrOCR for handwritten text recognition (HTR)htr-pipeline— full end-to-end HTR pipeline: text detection + recognition + output formatting
All training was performed in Kaggle Notebooks using free cloud GPUs.
| Model | Notebook | Task |
|---|---|---|
| CRAFT | CRAFT: Russian Dialectic Text Detection | Text Detection |
| YOLO | YOLO: Russian Dialectic Text Detection | Text Detection |
| TrOCR | TrOCR: Russian Dialectic HTR with PEFT (DoRA) | Handwritten Text Recognition (HTR) |
Note: These notebooks are demonstration versions (final experiment configurations). The original development notebooks are private and will not be published.
Fine-tuned model checkpoints are available on Hugging Face:
| Model | Hugging Face Link |
|---|---|
| CRAFT | Not published due to low performance — weights available in Kaggle notebook output (see above). |
| YOLO11x-dialectic | huggingface.co/Daniil-Domino/yolo11x-dialectic |
| TrOCR-base-Ru-dialectic | huggingface.co/Daniil-Domino/trocr-base-ru-dialectic |
The following datasets were used and are available on Hugging Face:
| Dataset | Hugging Face Link | Task |
|---|---|---|
| school_notebooks_RU | huggingface.co/datasets/ai-forever/school_notebooks_RU | Text Detection |
| Ru-dialectological-stackmix | huggingface.co/datasets/CherryJam/ru-dialectological-stackmix | Handwritten Text Recognition (HTR) |
| Ru-dialectological-fonts-aug | huggingface.co/datasets/CherryJam/ru-dialectological-fonts-aug | Handwritten Text Recognition (HTR) |
If you're running Kaggle notebooks listed above, all required dependencies are already included in the environment.
For local or custom environments, install the dependencies manually using:
pip install -r requirements.txt --quietTested with Python version 3.10.18
Matvienko Daniil - Machine Learning
Starikov Alexander - Dataset Building, Backend