ManchuAI-OCR

Optical Character Recognition for Manchu script using multiple model architectures.

Experimental Environment

• CPU: Intel Core i9-13900KS (32 cores)
• GPU: NVIDIA RTX 6000 Ada Generation (49GB VRAM)
• RAM: 188GB

Installation

uv sync

Quick Start (For Llama-3.2-11B)

1. Download Data

python scripts/download_data.py

2. Train Llama-3.2-11B Model

python scripts/finetune_llama32_11b.py

Train Llama-3.2-11B model on Manchu OCR datasets.

3. Evaluate Llama-3.2-11B Model

python scripts/infer_llama32_11b.py

Quick Start (For All Models)

Quick Start

The scripts/ folder contains the main entry points:

1. Train Models

python scripts/train.py

Trains VLM and CRNN models on Manchu OCR datasets.

2. Evaluate Models

python scripts/evaluate.py

Evaluates trained models on validation and test datasets.

3. Generate Figures

python scripts/generate_figures.py

Creates performance comparison charts and analysis figures.

Models

Vision Language Models

• qwen-25-3b/7b: Qwen2.5-VL-3B/7B
• llama-32-11b: Llama-3.2-11B

CRNN Models

• crnn-base-3m: Convolutional Recurrent Neural Network

Closed Domain Models

• openai-41: OpenAI GPT-4.1-2025-04-14

Results

Results are saved in results/ directory.