ai4bharat/Indic-Rag-Suite

🌏 Multilingual Indic RAG Suite

A comprehensive multilingual Q&A dataset containing 19 languages, focusing on Indian languages with English support. This dataset is optimized for RAG (Retrieval-Augmented Generation) applications, question-answering systems, and multilingual NLP research.

📊 Dataset Overview

Language Code	Language Name	Data File
as	Assamese	Assamese_Unfiltered.jsonl
bn	Bengali	Bengali_Unfiltered.jsonl
en	English	English_Unfiltered.jsonl
gu	Gujarati	Gujarati_Unfiltered.jsonl
hi	Hindi	Hindi_Unfiltered.jsonl
kn	Kannada	Kannada_Unfiltered.jsonl
ks	Kashmiri	Kashmiri_Unfiltered.jsonl
mai	Maithili	Maithili_Unfiltered.jsonl
ml	Malayalam	Malayalam_Unfiltered.jsonl
mni	Manipuri	Manipuri_Unfiltered.jsonl
mr	Marathi	Marathi_Unfiltered.jsonl
ne	Nepali	Nepali_Unfiltered.jsonl
or	Odia	Odia_Unfiltered.jsonl
pa	Punjabi	Punjabi_Unfiltered.jsonl
sa	Sanskrit	Sanskrit_Unfiltered.jsonl
sat	Santali	Santali_Unfiltered.jsonl
ta	Tamil	Tamil_Unfiltered.jsonl
te	Telugu	Telugu_Unfiltered.jsonl
ur	Urdu	Urdu_Unfiltered.jsonl

🚀 Quick Start

from datasets import load_dataset

# Load specific language datasets
hindi_data = load_dataset("ai4bharat/Indic-Rag-Suite", "hi")
print(f"Hindi samples: {len(hindi_data['train'])}")

# Load Bengali data
bengali_data = load_dataset("ai4bharat/Indic-Rag-Suite", "bn")
print(f"Bengali samples: {len(bengali_data['train'])}")

# Load English data
english_data = load_dataset("ai4bharat/Indic-Rag-Suite", "en")
print(f"English samples: {len(english_data['train'])}")

🔧 Available Configurations

Each language has its own configuration:

as (Assamese), bn (Bengali), en (English), gu (Gujarati), hi (Hindi), kn (Kannada), ks (Kashmiri), mai (Maithili), ml (Malayalam), mni (Manipuri), mr (Marathi), ne (Nepali), or (Odia), pa (Punjabi), sa (Sanskrit), sat (Santali), ta (Tamil), te (Telugu), ur (Urdu)

📋 Dataset Structure

Each example contains:

{
    "text": "Combined Q&A text with context",           # Main text field for RAG
    "language": "hi",                                # ISO 639-1 language code
    "source_lang": "hi",                            # Original source language
    "title": "Page title from Wikipedia",           # Article title
    "wiki_id": "12345",                             # Wikipedia page ID
    "url": "https://hi.wikipedia.org/wiki/...",     # Source URL
    "question": "Original question",                # Individual question
    "answer": "Answer to the question",             # Individual answer
    "reasoning": "Reasoning behind the answer",     # Explanation/reasoning
    "paragraph": "Source paragraph context"         # Original context paragraph
}

💡 Usage Examples

Basic Loading

from datasets import load_dataset

# Load Hindi dataset
hindi_ds = load_dataset("ai4bharat/Indic-Rag-Suite", "hi")['train']
print(f"First Hindi sample: {hindi_ds[0]['text'][:100]}")

# Access individual components
sample = hindi_ds[0]
print(f"Question: {sample['question']}")
print(f"Answer: {sample['answer']}")
print(f"Reasoning: {sample['reasoning']}")

Multi-language Loading

# Load multiple languages
languages = ['hi', 'bn', 'ta', 'te', 'gu']
datasets = {}

for lang in languages:
    try:
        datasets[lang] = load_dataset("ai4bharat/Indic-Rag-Suite", lang)['train']
        print(f"{lang}: {len(datasets[lang]):,} samples")
    except Exception as e:
        print(f"Failed to load {lang}: {e}")

Question-Answering Usage

# Load specific language for Q&A
data = load_dataset("ai4bharat/Indic-Rag-Suite", "hi")['train']

# Analyze Q&A patterns
for i in range(3):
    sample = data[i]
    print(f"Q: {sample['question']}")
    print(f"A: {sample['answer']}")
    print(f"Context: {sample['paragraph'][:100]}...")
    print("-" * 50)

RAG Application

# Use for retrieval-augmented generation
from datasets import load_dataset

# Load dataset
dataset = load_dataset("ai4bharat/Indic-Rag-Suite", "hi")['train']

# Extract texts for embedding
documents = [sample['text'] for sample in dataset]
questions = [sample['question'] for sample in dataset]
contexts = [sample['paragraph'] for sample in dataset]

# Use with your embedding model
print(f"Ready for RAG: {len(documents)} documents")

Streaming for Large Datasets

# Memory-efficient streaming
from datasets import load_dataset

# Stream large datasets
dataset = load_dataset("ai4bharat/Indic-Rag-Suite", "en", streaming=True)
train_stream = dataset['train']

# Process in batches
for i, example in enumerate(train_stream):
    if i >= 10:  # Process first 10 examples
        break
    print(f"Example {i+1}: {example['question'][:50]}...")

🎯 Use Cases

• 🤖 RAG Applications: Retrieval-augmented generation for Indian languages
• ❓ Question Answering: Training multilingual QA systems
• 🧠 Language Model Training: Fine-tuning language models
• 🔄 Cross-lingual Transfer: Transfer learning research
• 📊 NLP Research: Comparative studies across languages
• 🏷️ Text Classification: Multi-class classification tasks
• 📚 Knowledge Extraction: Building knowledge bases
• 🔍 Information Retrieval: Training retrieval models

⚡ Performance Features

• ✅ Individual Configs: Load only the language you need
• ✅ JSONL Format: Fast loading and processing
• ✅ Streaming Support: Handle large datasets efficiently
• ✅ Rich Metadata: Multiple fields for different use cases
• ✅ Standardized Schema: Consistent across all languages
• ✅ Language Codes: Standard ISO 639-1 codes

🏷️ Language Reference

Code	Language	Family	Script	Region
as	Assamese	Indo-Aryan	Bengali	Northeast India
bn	Bengali	Indo-Aryan	Bengali	Bengal
en	English	Germanic	Latin	Global
gu	Gujarati	Indo-Aryan	Gujarati	Gujarat
hi	Hindi	Indo-Aryan	Devanagari	North India
kn	Kannada	Dravidian	Kannada	Karnataka
ks	Kashmiri	Indo-Aryan	Arabic/Devanagari	Kashmir
mai	Maithili	Indo-Aryan	Devanagari	Bihar/Nepal
ml	Malayalam	Dravidian	Malayalam	Kerala
mni	Manipuri	Sino-Tibetan	Meetei Mayek	Manipur
mr	Marathi	Indo-Aryan	Devanagari	Maharashtra
ne	Nepali	Indo-Aryan	Devanagari	Nepal
or	Odia	Indo-Aryan	Odia	Odisha
pa	Punjabi	Indo-Aryan	Gurmukhi	Punjab
sa	Sanskrit	Indo-Aryan	Devanagari	Ancient/Classical
sat	Santali	Austroasiatic	Ol Chiki	Tribal regions
ta	Tamil	Dravidian	Tamil	Tamil Nadu
te	Telugu	Dravidian	Telugu	Andhra Pradesh
ur	Urdu	Indo-Aryan	Arabic	Pakistan/India

📚 Citation

@dataset{indic_rag_suite_2024,
    title={Multilingual Indic RAG Suite},
    author={AI4Bharat},
    year={2024},
    publisher={Hugging Face},
    howpublished={\url{https://huggingface.co/datasets/ai4bharat/Indic-Rag-Suite}},
    note={A comprehensive multilingual Q&A dataset for Indian languages}
}

📄 License

This dataset is released under the MIT License.

🤝 Contributing

We welcome contributions! Please open an issue for:

• Data quality improvements
• Additional languages
• Feature requests
• Bug reports

---

Built with ❤️ for advancing multilingual NLP in Indian languages