nexuss0781/GeezOCR

Frontier Amharic and Ge'ez OCR engineering, training, evaluation, and deployment # GeezOCR GeezOCR is an open engineering project for building a high-quality, high-throughput OCR system for Amharic and Ge’ez documents. The project starts from Tesseract’s floating-point Amharic LSTM model and aims to improve recognition through verified document data, active learning, human-reviewed ground truth, reproducible fine-tuning sessions, C++ performance engineering, and rigorous evaluation. ## Project status **Current phase: Phase 1 — Project charter and measurable targets.** The immediate objective is to freeze the project scope, evaluation protocol, data policy, and model-release criteria before large-scale PDF processing or fine-tuning begins. ## Engineering principles GeezOCR will preserve original documents, treat AI output as annotation assistance rather than ground truth, require human verification for training labels, split evaluation by document rather than randomly by line, retain a locked test set, and record hashes for data, checkpoints, configurations, and released models. The official Tesseract C++ LSTM training and inference implementation will be used first. Custom C++ components will be added around it for deterministic PDF/page processing, line preparation, batch inference, evaluation, and throughput benchmarking. The float model will remain the quality and fine-tuning reference; an integer model will be evaluated separately for speed and memory efficiency. ## Planned engineering areas | Area | Directory | |---|---| | C++ OCR and performance tools | `cpp/` | | Training configurations and scripts | `training/` | | Dataset schemas and manifests | `data/` | | Model registry and release metadata | `models/` | | Evaluation and error analysis | `evaluation/` | | Project documentation | `docs/` | | Automated workflows | `.github/workflows/` | ## Roadmap 1. Freeze the project charter and measurable targets. 2. Establish a reproducible C++/Tesseract build. 3. Design data storage, metadata, and dataset versioning. 4. Build PDF ingest …