A curated list of NLP resources, tools, datasets, and models for the Mon language (ISO 639-3:
mnw).
Mon is spoken by roughly one million people across Myanmar and Thailand, and UNESCO classifies it as vulnerable. Open tooling for it has been scarce. This list collects the software, models, and data that exist for working with Mon (ISO 639-3: mnw) — starting with the open-source MonOCR ecosystem and growing to cover the wider field. Contributions are welcome.
- OCR & Text Recognition
- Language Models & Tools
- Corpora & Datasets
- Data Collection Tools
- Language Resources
- Research & Papers
- Related Projects & Community
Tools for turning images of Mon script into text. The projects below form a single open pipeline: mon_OCR trains the model, monocrhf distributes the weights, monocr-onnx and monocr run inference, and the MonOCR apps ship it to end users.
- MonOCR - Web, Android, and iOS apps that recognize Mon script fully on-device, with no data leaving the device. Try it at ocr.mondevhub.com or on Google Play; an iOS build is in App Store review.
- monocr (Hugging Face) - Pretrained MonOCR weights published in ONNX, Core ML, and PyTorch formats, trained on roughly 3M samples (CER 0.025, WER 0.211).
- monocr-onnx - Privacy-first, on-device OCR engine powered by ONNX Runtime, with unified cross-platform SDKs for Python (
monocr-onnx), JavaScript/Node.js (monocr), Go, and Rust. - monocr - Lightweight Python OCR package and
monocrCLI for Mon text, with on-device inference and weights pulled from the Hugging Face Hub. Install withpip install monocr.
- mon_OCR - Production training pipeline for Mon OCR using a MobileNetV3-Large backbone, 2-layer BiLSTM, and CTC head over a 315-character charset, with ONNX and TFLite export.
Generative models, language identification, and other NLP building blocks for Mon.
- mon-language-detector - Python library that classifies text as Mon, Burmese, or English using a fastText model, for server-side filtering or on-device use.
- mon-gpt-playground - Streamlit app for testing and comparing the Mon-LM language models (0.5B, 1.5B, and 3B) fine-tuned for Mon.
Text and image datasets for training and evaluating Mon models. Open Mon corpora are scarce; if you maintain or know of one, please see Contributing.
- MonCorpusCollection - Mon-language text corpus of around one million lines gathered from Wikipedia, news agencies, and social media for NLP research and model training.
- Mon Wikipedia - Mon-language edition of Wikipedia with several thousand articles, a widely used source of running Mon text for corpus building.
- mon.monnews.org - Mon-language news site that serves as a primary source of running Mon text; harvestable with the tools in Data Collection Tools.
- Mon e-book library - Collection of a few hundred digitized Mon-language books, mostly Buddhist and Pali texts, readable online as flipbooks.
- mon-corpus-scraper - Web scraping toolkit that collects Mon-language text from mon.monnews.org and other sources for corpus and NLP work, built on Playwright, BeautifulSoup, and Polars.
Scripts, fonts, encoding, and input tooling needed to handle Mon text correctly.
- Unicode Myanmar block (U+1000–U+109F) - Unicode chart for the Myanmar script, which encodes Mon along with its language-specific characters.
Academic and technical work on Mon NLP.
- Mon language (ISO 639-3: mnw) - ISO 639-3 reference entry for Mon, a useful starting point for language metadata.
- MonDevHub - Organization building open tools for the Mon language, including the MonOCR ecosystem.
- Mon Buddhist Foundation - UK-based charity promoting Mon Buddhist culture and preserving the Mon language through educational programs, courses, and cultural events.
Contributions are welcome! Read the contribution guidelines first, and please follow the Code of Conduct.