গড়েইপা লৌশিং (AI) | Bishnupriya Manipuri Language Development Project

Building open-source AI for 500,000+ Bishnupriya Manipuri speakers worldwide

Bishnupriya Manipuri (BPY) is spoken across Assam, Tripura, Manipur, and Bangladesh. Despite its rich cultural heritage, it has zero support in Google Translate, Microsoft Translator, or any major AI model.

We are changing that.

🎯 Our Mission

Develop and maintain open-source NLP resources for Bishnupriya Manipuri, including:

Machine Translation - English ↔ BPY
Language Models - LLM pretraining and fine-tuning
Training Datasets - Parallel corpora, monolingual text, evaluation sets
Tools & Research - Tokenizers, OCR, speech models

All resources are free, open-source, and community-driven under MIT/Apache 2.0 licenses.

🚀 Our Models

NLLB Bishnupriya Manipuri Series

Fine-tuned versions of Meta AI's NLLB-200 for English → BPY translation.

Model	Status	Notes
nllb-bpy-beng-v8-5-3-merged	✅ Production	Latest stable. 95%+ accuracy. Fixes number+noun patterns. Live endpoint available.
nllb-bpy-beng-v8-5-3	Adapter	LoRA weights for fine-tuning. Requires base NLLB-200.
nllb-bpy-beng-v9-0	🔨 In Progress	Training on Wikipedia + scanned book corpus. Target: 10k+ pairs.

Live Demo: https://manipuri.com/articles/bpy.php#translator

Key Fixes in V8.5.3:

Fifty books → য়াংখেইহান লেরিক (not লেরিকহান লেরিকহান)
Uses ben_Beng token for proper Bengali script rendering
1000x weighted number+noun patterns eliminate repetition bugs

📊 Datasets

Coming Soon: BPY Training Data Repository

We are building the first comprehensive open dataset for Bishnupriya Manipuri:

Planned releases:

bpy-parallel-v1 - 10k+ English↔BPY sentence pairs from Wikipedia, books, community
bpy-monolingual-v1 - 100k+ BPY sentences for LM pretraining
bpy-eval-v1 - Standard test set for MT evaluation

Current sources:

BPY Wikipedia: 25,958 articles → extracting clean sentences
Scanned books: Community contributions, OCR pipeline ready
Community submissions: Accepting parallel text at [contact link]

Want to contribute data? See (CONTRIBUTING) or email us.

🤝 Collaborate With Us

We welcome researchers, developers, and BPY speakers to join:

We need help with:

Data Collection - Scan books, transcribe text, translate sentences
Model Training - Fine-tune LLMs, experiment with architectures
Evaluation - Build test sets, human eval, error analysis
Tools - OCR for Bengali script, tokenizers, text normalization
Applications - Chatbots, TTS, ASR, educational tools

Tech Stack: PyTorch, Transformers, PEFT/LoRA, Hugging Face Hub, PHP/Python

Join the Community

GitHub: [Coming soon]
Discord: [Coming soon]
Email: usangraha@gmail.com
Website: manipuri.com - Bishnupriya Manipuri Dictionary Project

📚 Resources & Research

Base Models Used:

facebook/nllb-200-distilled-600M - Meta AI's multilingual MT
LoRA fine-tuning for parameter-efficient adaptation

Papers & Docs:

NLLB: No Language Left Behind [Meta AI, 2022]
Fine-tuning methodology: [Technical report coming soon]

Language Info:

ISO 639-3: bpy
Script: Bengali (ben_Beng)
Speakers: ~500,000
Regions: Assam, Tripura, Manipur (India), Bangladesh

📜 License & Citation

All models and datasets released under MIT License - free for commercial use.

If you use our work, please cite:

@misc{bishnupriya-manipuri-nllb-2026,
  title={NLLB Bishnupriya Manipuri: Open-Source Machine Translation},
  author={Bishnupriya Manipuri Language Development Project},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/BishnupriyaManipuri}
}

থাকাত | Thank you for supporting low-resource language AI.

গড়েইপা লৌশিং (AI) - Let's build AI together.