📚 ViuTranslate-Data

A 100% Authentic Human-Curated Parallel Corpus for English ↔ Hindi Neural Translation

🌐 ViuAI Studio | 🤖 Model Hub | 📊 Data Provenance | 💡 Python Usage

---

📌 Introduction & Overview

ViuTranslate-Data is a production-grade bilingual parallel corpus curated specifically for fine-tuning neural machine translation foundation models (like ViuAI Sarus-500M).

Every single sentence pair is extracted from official, published academic and institutional corpora. Zero synthetic data, zero LLM-prompted dialogues, and zero templated variations were used.

---

📊 Corpus Provenance & Split

Source Corpus	Institution	Curated Pairs	Description
IIT Bombay English-Hindi Corpus	CFILT, IIT Bombay	50,000 Pairs	Academic gold standard covering news, judicial, literature, and governmental texts
AI4Bharat Samanantar	IIT Madras / AI4Bharat	50,000 Pairs	Verified web & publication Indian language corpus
IIT Bombay Benchmark Test Set	CFILT, IIT Bombay	2,502 Pairs	Internationally accepted gold validation test suite
Total Corpus	—	102,502 Pairs	13,633,518 Active Training Tokens

---

🛡️ Rigorous Quality Control Guardrails

To prevent dataset contamination, hallucinations, and vocabulary pollution, all samples were audited through strict mathematical heuristics:

1. Length Ratio Boundary: $$\quad 0.40 \le \frac{\text{len}(\text{English})}{\text{len}(\text{Hindi})} \le 2.40$$ Pairs violating this ratio were strictly pruned to eliminate incomplete or runaway translations.
2. Script Purity & Unicode Verification:
   • English side enforced $\ge 50%$ Latin alphabetic characters ($[A-Za-z]$).
   • Hindi side enforced $\ge 40%$ Devanagari Unicode characters ($[\u0900-\u097F]$).
3. Hygiene & Sanitation:
   • 100% stripped of HTML tags, XML nodes, source code snippets, URLs, and file paths.
4. Exact Cryptographic Deduplication:
   • SHA-256 hash deduplication ensuring zero repeated pairs across train and validation sets.

---

📁 Repository Structure

ViuTranslate-Data/
├── README.md                          # Official Dataset Card
├── metadata.json                      # Dataset Configuration & Token Counts
│
├── raw/                               # Raw Curated Sentence Pairs (JSON)
│   ├── viu_translate_100k_train.json  # 100,000 Verified Academic Pairs (29.6 MB)
│   └── viu_translate_val.json         # 2,502 Gold Benchmark Pairs (1.18 MB)
│
└── Pre-tokenized Shards (.npy)        # Ready-to-Train Memory-Mapped Arrays
    ├── train_tokens.npy               # Token IDs (int32, 54.5 MB)
    ├── train_labels.npy               # Loss-Masked Target Tokens (int32, 54.5 MB)
    ├── train_offsets.npy              # Sample Boundaries (int64, 1.6 MB)
    ├── train_domains.npy              # Direction Flags (int32, 0.8 MB)
    ├── val_tokens.npy                 # Validation Tokens (int32, 2.2 MB)
    ├── val_labels.npy                 # Validation Labels (int32, 2.2 MB)
    ├── val_offsets.npy                # Validation Offsets (int64, 0.04 MB)
    └── val_domains.npy                # Validation Domains (int32, 0.02 MB)

---

💡 Loading the Dataset

1. Using Hugging Face datasets

from datasets import load_dataset

dataset = load_dataset("ViuAI/ViuTranslate-Data", data_files={"train": "raw/viu_translate_100k_train.json"})
print("Sample 0:", dataset["train"][0])

2. Loading Pre-tokenized Memory-Mapped Shards (Fastest for Training)

import numpy as np
from huggingface_hub import hf_hub_download

tokens_file = hf_hub_download(repo_id="ViuAI/ViuTranslate-Data", filename="train_tokens.npy", repo_type="dataset")
tokens = np.load(tokens_file, mmap_mode="r")
print(f"Loaded {len(tokens):,} memory-mapped tokens.")

---

📜 Citation

@misc{viutranslate_data2026,
  author = {ViuAI Research Team},
  title = {ViuTranslate-Data: A 100% Authentic Human-Curated Parallel Corpus for English-Hindi Translation},
  year = {2026},
  publisher = {Hugging Face},
  journal = {Hugging Face Dataset Hub},
  howpublished = {\url{https://huggingface.co/datasets/ViuAI/ViuTranslate-Data}}
}