🔍 IndicMSMARCO: Multilingual Information Retrieval Benchmark

A comprehensive multilingual variant of MS MARCO specifically tailored for Indian languages, featuring carefully selected queries and corresponding passages with high-quality translations.

🚀 Quick Start - Load Individual Languages

from datasets import load_dataset

# Load ONLY Hindi data (fast and efficient!)
hindi_data = load_dataset("ai4bharat/IndicMSMARCO", "hi")
print(f"Hindi queries: {len(hindi_data['train'])} samples")

# Load ONLY Bengali data  
bengali_data = load_dataset("ai4bharat/IndicMSMARCO", "bn")
print(f"Bengali queries: {len(bengali_data['train'])} samples")

# Access query-passage pairs
for example in hindi_data['train'][:3]:
    print(f"Query: {example['query']}")
    print(f"Passage: {example['passage'][:200]}...")
    print("---")

📊 Dataset Overview

• Total Samples: 12,999
• Languages: 13 languages
• Source: MS MARCO development set
• Quality: Human-verified translations
• Task: Information Retrieval / Passage Ranking

🎯 Key Features

• Topic Diversity: Science, history, politics, health, technology
• Query Complexity: Simple factual, descriptive, and complex entity-based queries
• Balanced Representation: Short, medium, and long-form queries
• High-Quality Translations: Professional translation and verification
• Consistent Structure: Normalized schema across all languages

📋 Available Languages (13 total)

Code	Language	Load Command	Sample Count
as	Assamese	load_dataset('ai4bharat/IndicMSMARCO', 'as')	~999
bn	Bengali	load_dataset('ai4bharat/IndicMSMARCO', 'bn')	~999
gu	Gujarati	load_dataset('ai4bharat/IndicMSMARCO', 'gu')	~999
hi	Hindi	load_dataset('ai4bharat/IndicMSMARCO', 'hi')	~999
kn	Kannada	load_dataset('ai4bharat/IndicMSMARCO', 'kn')	~999
ml	Malayalam	load_dataset('ai4bharat/IndicMSMARCO', 'ml')	~999
mr	Marathi	load_dataset('ai4bharat/IndicMSMARCO', 'mr')	~999
ne	Nepali	load_dataset('ai4bharat/IndicMSMARCO', 'ne')	~999
or	Odia	load_dataset('ai4bharat/IndicMSMARCO', 'or')	~999
pa	Punjabi	load_dataset('ai4bharat/IndicMSMARCO', 'pa')	~999
ta	Tamil	load_dataset('ai4bharat/IndicMSMARCO', 'ta')	~999
te	Telugu	load_dataset('ai4bharat/IndicMSMARCO', 'te')	~999
ur	Urdu	load_dataset('ai4bharat/IndicMSMARCO', 'ur')	~999

💡 Usage Examples

Information Retrieval Evaluation

from datasets import load_dataset

# Load Hindi benchmark
dataset = load_dataset("ai4bharat/IndicMSMARCO", "hi")
queries = dataset['train']

# Extract queries and passages for retrieval evaluation
for item in queries:
    query_id = item['query_id']
    query_text = item['query']
    passage_text = item['passage']
    relevance = item['relevance_score']
    
    # Use for your retrieval model evaluation
    print(f"Query {query_id}: {query_text}")
    print(f"Relevant passage: {passage_text[:100]}...")

Cross-lingual Retrieval Benchmark

# Compare retrieval across languages
languages = ['as', 'bn', 'gu', 'hi']
results = {}

for lang in languages:
    dataset = load_dataset("ai4bharat/IndicMSMARCO", lang)
    results[lang] = dataset['train']
    print(f"{lang}: {len(results[lang])} query-passage pairs")

# Evaluate your multilingual retrieval model
for lang_code, queries in results.items():
    # Run your retrieval evaluation here
    pass

📋 Dataset Structure

{
    "query_id": "1234567",
    "query": "भारत की राजधानी क्या है?",
    "passage": "भारत की राजधानी नई दिल्ली है। यह देश के उत्तरी भाग में स्थित है...",
    "passage_id": "7654321",
    "language": "hi",
    "answer": "नई दिल्ली",
    "title": "भारत की राजधानी",
    "query_type": "factual",
    "relevance_score": 1.0,
    "is_selected": true,
    "text": "Query: भारत की राजधानी क्या है? | Passage: भारत की राजधानी नई दिल्ली है...",
    "dataset": "IndicMSMARCO",
    "source": "MS MARCO translated to Indian languages",
    "meta": "{\"model\": \"translation_model\", \"verified\": true}"
}

⚡ Performance & Loading Tips

• Single Language Loading: Always use config name for fastest loading
• Streaming: Use streaming=True for memory-efficient processing
• Batch Evaluation: Load full train split for comprehensive benchmarking
• Cross-lingual: Compare same query_id across languages

🎯 Use Cases

• 🔍 Information Retrieval: Benchmark multilingual retrieval systems
• 🤖 RAG Evaluation: Test retrieval-augmented generation systems
• 📊 Cross-lingual IR: Evaluate cross-language information retrieval
• 🧪 Model Comparison: Compare multilingual embedding models
• 📚 Academic Research: Multilingual IR and NLP research

📖 Citation

If you use IndicMSMARCO in your research, please cite:

@dataset{indic_msmarco_2024,
  title={IndicRAGSuite: LargeScale Datasets and a Benchmark for Indian Language RAG Systems},
  author={Pasunuti Prasanjith,Prathmesh B More,Anoop Kunchukuttan, Raj Dabre},
  year={2025},
  {journal = {arXiv preprint arXiv:2506.01615},
  url={https://huggingface.co/datasets/ai4bharat/IndicMSMARCO}
}

🔧 Technical Details

• Format: JSONL files per language
• Encoding: UTF-8
• Schema: Normalized MS MARCO structure
• Quality Control: Multi-stage validation process

---

Built for multilingual information retrieval • Human-verified quality • Ready for benchmarking