TuneOCR

TuneOCR is a stable, production-ready, open-source framework for fine-tuning OCR models. It helps researchers, developers, and hobbyists train and adapt small OCR models more effectively across diverse document types and languages while keeping compute and cost practical.

---

Overview

TuneOCR provides reproducible training pipelines, dataset utilities, and helpers for flexible token / special-token positioning so you can experiment with language tags, transcription prompts, and label-formatting strategies without modifying model internals.

Key capabilities:

• Fine-tune a variety of compact OCR and document-understanding models.
• Reformat or inject special tokens (e.g., language tags) in training targets.
• Simple data collators and training scripts ready for CPU/GPU or small cloud instances.
• Extensible plugin/back-end system for new model integrations.
• Built-in deterministic runs and logging to reproduce experiments.

---

Supported Architectures

TuneOCR ships with integrations or examples for the following architectures:

1. TrOCR — Single-line handwriting and printed text.
2. Donut — Document-to-JSON extraction (receipts, invoices, structured forms).
3. VLMs — Visual-linguistic models for complex document QA and understanding.
4. QWEN OCR — Multi-language, high-capacity OCR models.
5. Nanonets OCR — Lightweight, fast OCR suitable for scanned forms and field capture.
6. OlmOCR — Flexible research-oriented OCR framework for experimentation.

If a backend you need is missing, TuneOCR is designed so you can add one quickly (processor, model wrapper, data prep).

---

Why TuneOCR?

• Open source & community-first — Inspect, reuse, and extend the training recipes.
• Small-model focused — Emphasizes parameter-efficient approaches (LoRA/adapters, selective fine-tuning) and practical defaults so you can iterate on modest hardware.
• Flexible label formatting — Move language/task tokens (e.g., <|en|>, <|transcribe|>) to start, middle, or end of targets to test what works for your data.
• Reproducible experiments — Deterministic seeds, config-driven runs, and logging for fair comparisons.
• Practical evaluations — WER / CER and structured extraction checks (for Donut-like workflows).

---

Installation

Python 3.11 recommended

Install with Pip

pip install -r requirements.txt

Contribution & Community

Contributors are super welcome! Help the project grow by opening issues and submitting PRs.

How to contribute:

Open an issue for bugs, feature requests, or new backend proposals. Include:

• Problem statement / feature description

• Minimal repro or sample dataset (if applicable)

• Expected behavior or desired API

Create a Pull Request (PR) for fixes, features, docs, or new backends. PR checklist we appreciate:

• Descriptive title and summary of the change

• Tests where appropriate (unit or small integration)

• Updated CHANGELOG.md and documentation for visible changes

Use feature branches named like feature/qwen-backend or fix/token-positioning-bug.

Contributors

All contributors are welcome. To be acknowledged:

• Add yourself to CONTRIBUTORS.md via PR, or open an issue to request inclusion.

• Major contributions will be noted in CHANGELOG.md and release notes.

Contact / Maintainer

• Author / Maintainer: Emkay Nguyen
• Email: minhkhoinguyendo1210@gmail.com

For coordination, partnership, or academic collaboration, open an issue or email the maintainer.