README.md

metadata

language:
  - ctd
  - zom
  - en
license: apache-2.0
task_categories:
  - translation
  - text-generation
  - fill-mask
  - sentence-similarity
  - token-classification
tags:
  - zomi
  - chin
  - kuki-mizo
  - low-resource
  - nllb
  - neural-machine-translation
  - ai-for-good
  - culture-preservation
pretty_name: Zomi-English Parallel Intelligence Corpus (ZEPIC)
size_categories:
  - 10K<n<100K
config_names:
  - default
dataset_info:
  features:
    - name: english
      dtype: string
    - name: zomi
      dtype: string
    - name: source
      dtype: string
    - name: domain
      dtype: string
  splits:
    - name: train
      num_bytes: 5242880
      num_examples: 15000
    - name: test
      num_bytes: 524288
      num_examples: 1500
  download_size: 2500000
  dataset_size: 5767168

🌏 ZOMI OPEN AI INITIATIVE 🌏

Bridging the Digital Divide via Neural Machine Translation

🌐 Live Demo | 💬 Join Community | 📖 Developer Docs

---

📑 Table of Contents

1. Project Overview & Mission
2. Dataset Statistics & Analytics
3. The Zomi AI Architecture
4. Linguistic Profile: Understanding Zomi
5. Advanced Usage Guide (Python/API)
6. Model Benchmarks & Performance
7. Contribution & Roadmap
8. Data Governance, Ethics & License
9. Citation

---

🚀 Project Overview & Mission

The Zomi Language (ISO 639-3: ctd / zom) is the lingua franca of the Zomi people across the highlands of Myanmar, India (Manipur/Mizoram), and Bangladesh. Despite a vibrant literary history and millions of speakers, it remains "Low-Resource" in the global digital ecosystem.

The Zomi-English Parallel Intelligence Corpus (ZEPIC) is the world's first open-source, machine-readable dataset designed specifically for Large Language Model (LLM) fine-tuning. It moves beyond simple dictionary matching to capture semantic context, idiomatic expressions, and syntactic alignment.

Core Objectives

• Digital Sovereignty: Owning our data to prevent digital extinction and ensuring the Zomi identity thrives in the AI era.
• Educational Access: Enabling instant translation of global knowledge (Science, Medicine, Law) into Zomi.
• Cultural Archiving: Preserving oral history and literature through advanced text generation.

---

📊 Dataset Statistics & Analytics

We employ rigorous data cleaning (deduplication, length-ratio filtering) to ensure high-quality training signals.

Metric	Value	Details
Total Sentence Pairs	15,000+	Validated English-Zomi bitexts.
Token Count (En)	~250,000	English Source Tokens.
Token Count (Zomi)	~240,000	Zomi Target Tokens (Agglutinative).
Vocabulary Size	28,400	Unique Zomi word forms.
Avg. Sentence Length	14.2 Words	Ideal for Transformer context windows.
Dialect Standard	Zomi (Standard)	Based on the Tedim literary standard.

Domain Distribution

pie title Source Domain Distribution
    "General Conversation" : 40
    "Religious/Theological" : 30
    "News & Media" : 15
    "Educational/Science" : 10
    "Legal/Formal" : 5