The Merged Nepali Health FAQ Dataset is a large-scale, high-quality multilingual health information resource containing 3,174 question-answer pairs in Nepali language. This dataset is specifically designed for training conversational AI models, information retrieval systems, and question-answering applications focused on health and medical topics.

Key Statistics

Metric	Value
Total Records	3,174
Language	Nepali (नेपाली)
Script	Devanagari (देवनागरी)
License	Apache-2.0 (Permissive)
Data Format	JSONL (JSON Lines)
Task Type	Instruction-Following / Question-Answering
Generation Type	Real (Authentic)
Unique Health Domains	3 Major + 20+ Sub-domains
Unique Sources	20+ Medical Organizations & Healthcare Providers
Average Question Length	7.6 words
Average Answer Length	39.9 words

1. BEHAVIOR DIVERSITY ANALYSIS

The dataset includes diverse response behaviors designed to handle different types of health inquiries with varying levels of detail and formality.

Behavior Distribution

Behavior Type	Count	Percentage	Description
जानकारीमूलक उत्तर दिने (Informative Answer)	1,458	45.9%	Provides comprehensive, factual health information while maintaining the core meaning and facts of the original answer translated to Nepali
सामान्य जानकारी दिने (General Information)	1,363	42.9%	Delivers general health knowledge without excessive technical details
छोटो उत्तर माग्ने (Brief Answer Request)	215	6.8%	Concise responses addressing specific health queries
प्रश्नको प्रत्यक्ष उत्तर दिने (Direct Answer)	100	3.2%	Direct, straightforward responses to health questions
छोटो र जानकारीमूलक उत्तर दिने (Short & Informative)	38	1.2%	Balanced approach combining brevity with informative content

Behavior Insights

Informative vs. General Split: Nearly 89% of responses adopt an informative or general knowledge approach, ensuring users receive substantive health information.
Response Conciseness: About 7% of entries employ brief responses, providing quick answers for straightforward health queries.
Direct Answers: Only 3.2% use purely direct answers, indicating the dataset prioritizes comprehensive information over minimal responses.
Mixed Approaches: A small subset (1.2%) combines brevity with information, offering a balanced middle ground for moderately complex questions.

Use Cases

Training conversational agents that provide medically accurate information
Building health chatbots with varying response styles
Developing QA systems that adapt response length to question complexity
Content generation for health education platforms

2. QUESTION PATTERN DIVERSITY ANALYSIS

The dataset encompasses multiple types of health-related questions, each serving different information needs.

Question Type Distribution

Question Type	Count	Percentage	Example Patterns
तथ्यमा आधारित प्रश्न (Fact-Based Questions)	2,831	89.2%	"आधारभूत चयापचय दर किन महत्त्वपूर्ण छ?" (Why is basal metabolic rate important?)
प्रश्नोत्तर प्रश्न (Interrogative Questions)	101	3.2%	Questions phrased as inquiries requiring verification
स्वास्थ्यसम्बन्धी प्रश्न (Health-Specific Questions)	100	3.2%	Specialized health domain questions
प्रश्नोत्तर (Q&A Format)	100	3.2%	Structured question-answer pairs
सामान्य प्रश्न (General Questions)	4	0.1%	Broad health inquiries
N/A (Unclassified)	38	1.2%	Uncategorized entries

Question Pattern Characteristics

Fact-Based Questions (89.2%)

Characteristics: Seek objective health information
Typical Format: "What is...?", "Why does...?", "How does...?"
Examples:
- "परागकणको स्तर भनेको के हो?" (What is pollen count?)
- "समयसँगै मेरो आधारभूत चयापचय दर परिवर्तन हुन सक्छ?" (Can my basal metabolic rate change over time?)
- "दमले मेरो जीवन प्रत्याशामा असर गर्छ?" (Does asthma affect my life expectancy?)

Interrogative Questions (3.2%)

Characteristics: Phrased as inquiries requiring yes/no or verification answers
Format: Direct questioning of medical concepts
Use Case: Training systems for validation-based QA

Health-Specific Questions (3.2%)

Characteristics: Domain-specific health terminology
Examples: "कोलोरेक्टल क्यान्सरको चरण कस्तो निर्धारण गरिन्छ?" (How is colorectal cancer staged?)

Structured Q&A (3.2%)

Characteristics: Pre-formatted question-answer pairs
Use Case: Direct fact retrieval systems

Question Complexity Metrics

Metric	Value	Range
Average Word Count	7.6 words	2-28 words
Minimum Length	2 words	Simple queries
Maximum Length	28 words	Complex multi-part questions
Median Length	~6 words	Typical query

Question Pattern Examples by Complexity

Simple (2-4 words)

"के असथेमा घातक हुन्छ?" (Is asthma fatal?)
"कोलोन क्यान्सर के हो?" (What is colon cancer?)

Moderate (5-10 words)

"मेरो खानपानले आधारभूत चयापचय दरलाई असर गर्छ?" (Does my diet affect basal metabolic rate?)
"उर्वर समयावधि भनेको के हो?" (What is the fertile window?)

Complex (15+ words)

"तपाईंका लागि घाँसको परागबाट हुने ज्वरोमा प्रयोग गर्ने कुन प्रकारको नाकको छर्काइ उपयुक्त हुन्छ?" (What type of nasal spray is suitable for hay fever treatment for you?)
"क्या व्यक्तिगत स्वतन्त्रता भत्ता र अपाङ्गता जीवनयापन भत्ताबीच के फरक छ?" (What is the difference between personal independence payment and disability living allowance?)

3. DOMAIN AND CATEGORY DIVERSITY

Primary Domain Distribution

Domain	Count	Percentage	Description
स्वास्थ्य (Health)	3,036	95.7%	Core health and medical topics
एस्थेटिक मेडिसिन (Aesthetic Medicine)	100	3.2%	Cosmetic and aesthetic procedures
शिक्षा (Education)	38	1.2%	Educational/academic health topics

Category Distribution (Top 10)

Category	Count	Percentage	Focus Area
स्वास्थ्य सम्बन्धी प्रश्नोत्तर (Health Q&A)	1,235	38.9%	General health Q&A
स्वास्थ्यसम्बन्धी प्रायः सोधिने प्रश्न (Frequently Asked Health)	938	29.6%	Common health questions
क्यान्सर सम्बन्धी प्रायः सोधिने प्रश्न (Cancer FAQ)	191	6.0%	Cancer-specific questions
एस्थेटिक अस्पताल प्रश्नोत्तर (Aesthetic Hospital Q&A)	101	3.2%	Cosmetic procedures
छाती तथा एलर्जीसम्बन्धी स्वास्थ्य प्रश्न (Chest & Allergy Health)	100	3.2%	Respiratory and allergy issues
स्वास्थ्य, फिटनेस र पोषण (Health, Fitness & Nutrition)	100	3.2%	Lifestyle and wellness
भिटिलिगो तथा छालाको स्वास्थ्य (Vitiligo & Skin Health)	99	3.1%	Dermatological conditions
अज्ञात (Unknown)	99	3.1%	Unclassified entries
राष्ट्रिय स्वास्थ्य सेवा (NHS) जागिर (NHS Jobs FAQ)	97	3.1%	Healthcare employment
गर्भनिरोधक सम्बन्धी प्रायः सोधिने प्रश्न (Contraception FAQ)	53	1.7%	Family planning

Comprehensive Category Breakdown

Category	Count	%	Sub-topics
Health Q&A	1,235	38.9%	Radiology, imaging, general health
Frequently Asked	938	29.6%	Patient.info compilations
Cancer-Related	191	6.0%	Colorectal cancer focus
Aesthetic	101	3.2%	Cosmetic procedures
Respiratory	100	3.2%	Asthma, allergies, respiratory care
Fitness & Nutrition	100	3.2%	Diet, exercise, weight management
Dermatology	99	3.1%	Vitiligo, skin conditions
Healthcare Jobs	97	3.1%	NHS employment, careers
Reproductive Health	53	1.7%	Contraception, fertility
Medical Education	38	1.2%	MBBS in Nepal
Insurance	29	0.9%	Health insurance information
Blood Donation	26	0.8%	Blood donation FAQs
Ophthalmology	13	0.4%	Eye health
Other Specialties	90	2.8%	Organ transplant, HIV/AIDS, mountain medicine

Sub-Domain Diversity (Top 15)

Sub-Domain	Count	Percentage	Medical Specialty
रेडियोलोजी तथा इमेजिङ (Radiology & Imaging)	1,235	38.9%	Diagnostic imaging
विविध स्वास्थ्यसम्बन्धी विषय (Miscellaneous Health)	496	15.6%	General health topics
कोलोरेक्टल क्यान्सर (Colorectal Cancer)	191	6.0%	Oncology
हेल्दी च्वाइस एस्थेटिक (Healthy Choice Aesthetic)	101	3.2%	Cosmetic medicine
दम तथा एलर्जी (Asthma & Allergy)	100	3.2%	Pulmonology
बारम्बार सोधिने प्रश्नहरू (General FAQs)	100	3.2%	General knowledge
सामान्य स्वास्थ्य र व्यायाम (General Health & Exercise)	100	3.2%	Preventive medicine
भिटिलिगो सम्बन्धी (Vitiligo)	99	3.1%	Dermatology
राष्ट्रिय स्वास्थ्य सेवा जागिर (NHS Jobs)	97	3.1%	Healthcare careers
विविध स्वास्थ्य अवस्था (Miscellaneous Conditions)	88	2.8%	Various conditions
गर्भनिरोधक (Contraception)	53	1.7%	Reproductive health
नेपालमा एमबीबीएस (MBBS in Nepal)	38	1.2%	Medical education
स्वास्थ्य बीमा (Health Insurance)	29	0.9%	Insurance information
रक्तदान (Blood Donation)	26	0.8%	Hematology
आँखा अस्पताल (Eye Hospital)	13	0.4%	Ophthalmology

4. SOURCE AND ORGANIZATION DIVERSITY

The dataset is compiled from 20+ diverse health organizations and medical institutions, providing varied perspectives and expertise.

Source Distribution (Complete List)

Source Organization	Count	Percentage	Organization Type	Focus
RadFAQs	1,235	38.9%	Radiology Center	Diagnostic imaging Q&A
Patient.info	938	29.6%	Healthcare Portal	General patient information
Cancer Clinic Pune	191	6.0%	Oncology Clinic	Cancer treatment & management
CohereLabs/aya_dataset	101	3.2%	Research/Dataset	Multilingual health content
Chest Clinic of Kathmandu & Allergy Center	100	3.2%	Specialty Clinic	Respiratory & allergies
Harley Street Institute	100	3.2%	Medical Institute	General medical information
SAC Nepal	100	3.2%	Healthcare Provider	Nepalese health services
Fight Vitiligo	99	3.1%	Specialist Organization	Vitiligo/skin conditions
Global Healthcare Jobs	97	3.1%	Employment Portal	Healthcare careers
Nepal CRS Company	53	1.7%	Healthcare Company	Reproductive health
Future Education - MBBS in Nepal	38	1.2%	Educational Provider	Medical education
स्वास्थ्य बीमा बोर्ड (Health Insurance Board)	29	0.9%	Government Agency	Health insurance
राष्ट्रिय रक्तसञ्चार सेवा केन्द्र (National Blood Transfusion Service)	26	0.8%	Government Service	Blood donation & transfusion
Mechi Eye Hospital	13	0.4%	Eye Care Center	Ophthalmology
Swastha Aviyan Nepal	11	0.3%	Health Movement	General wellness
Kirtipur Hospital FAQ	10	0.3%	Hospital	General hospital information
राष्ट्रिय जनस्वास्थ्य प्रयोगशाला (National Public Health Lab)	10	0.3%	Government Lab	Public health
राष्ट्रिय एड्स तथा यौन रोग नियन्त्रण केन्द्र (National AIDS & STI Center)	7	0.2%	Government Center	HIV/AIDS & STI
MMSN	6	0.2%	Organization	Mountain medicine
World Health Organization (WHO)	6	0.2%	International Body	Public health guidelines
स्वास्थ्य सेवा विभाग अनलाइन सीपीडी (Health Service Online CPD)	4	0.1%	Training Provider	Professional development

Organization Type Breakdown

Type	Count	Percentage	Examples
Healthcare Portals/Websites	2,173	68.5%	Patient.info, RadFAQs
Specialty Clinics	293	9.2%	Chest Clinic, Eye Hospital, Cancer Clinic
Government Agencies	73	2.3%	Blood Bank, Insurance Board, Public Health Lab
Educational Providers	38	1.2%	Medical education institutions
International Organizations	6	0.2%	World Health Organization
Research/Dataset Sources	101	3.2%	CohereLabs, academic datasets
Other Providers	491	15.4%	Healthcare companies, movements

Geographic Diversity

Nepal-based Sources: 600+ records (SAC Nepal, Chest Clinic of Kathmandu, Mechi Eye Hospital, Kirtipur Hospital, National services)
India-based Sources: 191+ records (Cancer Clinic Pune)
UK-based Sources: 2,100+ records (Patient.info, RadFAQs, Harley Street Institute, NHS)
International Sources: 183+ records (WHO, CohereLabs multilingual data)

5. TEXT LENGTH AND CONTENT ANALYSIS

Question Length Distribution

Metric	Value	Range	Interpretation
Average Length	7.6 words	2-28 words	Moderately concise queries
Minimum Length	2 words	Simple queries	Very short health inquiries
Maximum Length	28 words	Complex queries	Multi-part or detailed questions
Median Length	~6 words	Typical query	Most questions are brief

Distribution Breakdown:

Very Short (2-4 words): ~25% | "What is asthma?" type questions
Short (5-8 words): ~50% | Most common query length
Medium (9-14 words): ~20% | Moderately detailed questions
Long (15-28 words): ~5% | Complex multi-part questions

Answer Length Distribution

Metric	Value	Range	Interpretation
Average Length	39.9 words	1-234 words	Comprehensive responses
Minimum Length	1 word	Single-word answers	Rare minimal responses
Maximum Length	234 words	Detailed explanations	Maximum detail answers
Median Length	~35 words	Typical answer	Standard informative response

Distribution Breakdown:

Minimal (1-5 words): ~8% | Brief direct answers
Short (6-20 words): ~30% | Quick informative answers
Medium (21-50 words): ~40% | Standard comprehensive answers
Long (51-100 words): ~18% | Detailed explanations
Very Long (100+ words): ~4% | Extensive medical information

Content Depth Analysis

Answer Length Category	Count (Est.)	Use Case
Minimal (1-5 words)	~254	Quick fact retrieval, yes/no answers
Short (6-20 words)	~952	Brief health tips, quick explanations
Medium (21-50 words)	~1,270	Standard health information
Long (51-100 words)	~572	Detailed medical explanations
Very Long (100+ words)	~126	Comprehensive disease information

6. DATASET COMPOSITION SUMMARY TABLE

Complete Statistical Overview

Aspect	Value	Count	Percentage
Total Records	-	3,174	100%
Behavior Types	Informative	1,458	45.9%
	General Info	1,363	42.9%
	Brief Answer	215	6.8%
	Direct Answer	100	3.2%
	Mixed Approach	38	1.2%
Question Types	Fact-Based	2,831	89.2%
	Interrogative	101	3.2%
	Health-Specific	100	3.2%
	Structured Q&A	100	3.2%
	General	4	0.1%
	Unclassified	38	1.2%
Primary Domains	Health	3,036	95.7%
	Aesthetic Medicine	100	3.2%
	Education	38	1.2%
Sources	Unique Organizations	20+	-
	Radiology Focus	1,235	38.9%
	General Health	938	29.6%
	Specialty Care	939	29.6%
Text Metrics	Avg Question	7.6	words
	Avg Answer	39.9	words
	Max Question	28	words
	Max Answer	234	words

7. HOW TO LOAD AND USE THE DATASET

7.1 Prerequisites

Before loading the dataset, ensure you have Python 3.7+ and the following libraries installed:

pip install pandas jsonlines numpy

7.2 Basic Loading - Simple Approach

import json

def load_nepali_health_faq_simple():
    """
    Load the Nepali Health FAQ dataset in a simple, straightforward manner.
    """
    dataset = []
    
    with open('merged_nepali_health_faq_dataset.jsonl', 'r', encoding='utf-8') as f:
        for line in f:
            record = json.loads(line)
            dataset.append(record)
    
    print(f"✓ Loaded {len(dataset)} records")
    return dataset

# Load the dataset
data = load_nepali_health_faq_simple()

# Access first record
first_record = data[0]
print(f"ID: {first_record['id']}")
print(f"Question: {first_record['conversations'][0]['value']}")
print(f"Answer: {first_record['conversations'][1]['value']}")

Output:

✓ Loaded 3174 records
ID: nepali_health_faq_00001
Question: तौल व्यवस्थापनका लागि आधारभूत चयापचय दर किन महत्त्वपूर्ण छ?
Answer: आधारभूत चयापचय दर थाहा पाउँदा शरीरलाई सामान्य रूपमा काम गर्न आवश्यक पर्ने आधारभूत क्यालोरीको संख्या बुझ्न मद्दत हुन्छ...

7.3 Advanced Loading - Structured Approach

import json
from typing import List, Dict, Any
import pandas as pd

class NepaliHealthFAQDataset:
    """
    Advanced loader for the Nepali Health FAQ Dataset with built-in parsing and analysis.
    """
    
    def __init__(self, filepath: str):
        self.filepath = filepath
        self.records = []
        self.metadata = {
            'behaviors': [],
            'question_types': [],
            'domains': [],
            'sources': []
        }
    
    def load(self) -> List[Dict[str, Any]]:
        """
        Load all records from the JSONL file.
        """
        print("📖 Loading Nepali Health FAQ Dataset...")
        
        with open(self.filepath, 'r', encoding='utf-8') as f:
            for idx, line in enumerate(f, 1):
                try:
                    record = json.loads(line)
                    self.records.append(record)
                    
                    # Extract metadata
                    metadata_json = json.loads(record.get('metadata_json', '{}'))
                    self.metadata['behaviors'].append(metadata_json.get('behavior'))
                    self.metadata['question_types'].append(metadata_json.get('question_type'))
                    self.metadata['domains'].append(metadata_json.get('generation_domain'))
                    self.metadata['sources'].append(record.get('source'))
                    
                except json.JSONDecodeError as e:
                    print(f"⚠️  Error parsing line {idx}: {e}")
                    continue
        
        print(f"✓ Successfully loaded {len(self.records)} records\n")
        return self.records
    
    def get_record(self, index: int) -> Dict[str, Any]:
        """
        Get a specific record by index.
        """
        if index >= len(self.records):
            raise IndexError(f"Index {index} out of range. Dataset has {len(self.records)} records.")
        return self.records[index]
    
    def get_qa_pair(self, index: int) -> Dict[str, str]:
        """
        Extract question and answer from a record.
        """
        record = self.get_record(index)
        conversations = record.get('conversations', [])
        
        question = None
        answer = None
        
        for conv in conversations:
            if conv.get('from') == 'human':
                question = conv.get('value', '')
            elif conv.get('from') == 'gpt':
                answer = conv.get('value', '')
        
        return {
            'question': question,
            'answer': answer,
            'source': record.get('source'),
            'behavior': self._get_metadata_value(index, 'behavior')
        }
    
    def _get_metadata_value(self, index: int, key: str) -> str:
        """
        Get metadata value for a record.
        """
        record = self.records[index]
        metadata = json.loads(record.get('metadata_json', '{}'))
        return metadata.get(key, 'N/A')
    
    def get_by_behavior(self, behavior: str) -> List[Dict[str, Any]]:
        """
        Get all records with a specific behavior.
        """
        results = []
        for idx, record in enumerate(self.records):
            metadata = json.loads(record.get('metadata_json', '{}'))
            if metadata.get('behavior') == behavior:
                results.append(record)
        return results
    
    def get_by_source(self, source: str) -> List[Dict[str, Any]]:
        """
        Get all records from a specific source.
        """
        return [r for r in self.records if r.get('source') == source]
    
    def get_by_domain(self, domain: str) -> List[Dict[str, Any]]:
        """
        Get all records from a specific domain.
        """
        results = []
        for record in self.records:
            metadata = json.loads(record.get('metadata_json', '{}'))
            if metadata.get('generation_domain') == domain:
                results.append(record)
        return results
    
    def get_statistics(self) -> Dict[str, Any]:
        """
        Get comprehensive statistics about the dataset.
        """
        from collections import Counter
        
        behaviors = Counter([b for b in self.metadata['behaviors'] if b])
        question_types = Counter([q for q in self.metadata['question_types'] if q])
        sources = Counter([s for s in self.metadata['sources'] if s])
        
        return {
            'total_records': len(self.records),
            'unique_sources': len(set(self.metadata['sources'])),
            'unique_behaviors': len(set(self.metadata['behaviors'])),
            'unique_question_types': len(set(self.metadata['question_types'])),
            'behaviors': dict(behaviors.most_common()),
            'question_types': dict(question_types.most_common()),
            'sources': dict(sources.most_common(10))
        }
    
    def to_dataframe(self) -> pd.DataFrame:
        """
        Convert dataset to a pandas DataFrame for easy analysis.
        """
        data = []
        
        for record in self.records:
            conversations = record.get('conversations', [])
            metadata = json.loads(record.get('metadata_json', '{}'))
            
            question = None
            answer = None
            
            for conv in conversations:
                if conv.get('from') == 'human':
                    question = conv.get('value', '')
                elif conv.get('from') == 'gpt':
                    answer = conv.get('value', '')
            
            data.append({
                'id': record.get('id'),
                'question': question,
                'answer': answer,
                'source': record.get('source'),
                'domain': metadata.get('generation_domain'),
                'category': metadata.get('generation_category'),
                'behavior': metadata.get('behavior'),
                'question_type': metadata.get('question_type'),
                'question_length': len(question.split()) if question else 0,
                'answer_length': len(answer.split()) if answer else 0
            })
        
        return pd.DataFrame(data)

# Usage Example
if __name__ == "__main__":
    # Initialize and load dataset
    dataset = NepaliHealthFAQDataset('merged_nepali_health_faq_dataset.jsonl')
    records = dataset.load()
    
    # Get statistics
    stats = dataset.get_statistics()
    print("📊 Dataset Statistics:")
    print(f"Total Records: {stats['total_records']}")
    print(f"Unique Sources: {stats['unique_sources']}")
    print(f"Unique Behaviors: {stats['unique_behaviors']}")
    print(f"\nTop 5 Behaviors:")
    for behavior, count in list(stats['behaviors'].items())[:5]:
        print(f"  • {behavior}: {count}")
    
    # Get a specific Q&A pair
    qa = dataset.get_qa_pair(0)
    print(f"\n📝 Sample Q&A Pair:")
    print(f"Question: {qa['question']}")
    print(f"Answer: {qa['answer'][:100]}...")
    print(f"Source: {qa['source']}")
    
    # Get records by behavior
    informative_qa = dataset.get_by_behavior('जानकारीमूलक उत्तर दिने')
    print(f"\n✓ Found {len(informative_qa)} informative Q&A pairs")
    
    # Get records by source
    patient_info = dataset.get_by_source('Patient.info')
    print(f"✓ Found {len(patient_info)} records from Patient.info")
    
    # Convert to DataFrame
    df = dataset.to_dataframe()
    print(f"\n📋 DataFrame Shape: {df.shape}")
    print(f"Average Question Length: {df['question_length'].mean():.1f} words")
    print(f"Average Answer Length: {df['answer_length'].mean():.1f} words")
    
    # Display first few rows
    print("\n🔍 First 5 Records:")
    print(df[['id', 'domain', 'behavior', 'question_length', 'answer_length']].head())

Output:

📖 Loading Nepali Health FAQ Dataset...
✓ Successfully loaded 3174 records

📊 Dataset Statistics:
Total Records: 3174
Unique Sources: 20
Unique Behaviors: 5

Top 5 Behaviors:
  • जानकारीमूलक उत्तर दिने: 1458
  • सामान्य जानकारी दिने: 1363
  • छोटो उत्तर माग्ने: 215
  • प्रश्नको प्रत्यक्ष उत्तर दिने: 100
  • छोटो र जानकारीमूलक उत्तर दिने: 38

📝 Sample Q&A Pair:
Question: तौल व्यवस्थापनका लागि आधारभूत चयापचय दर किन महत्त्वपूर्ण छ?
Answer: आधारभूत चयापचय दर थाहा पाउँदा शरीरलाई सामान्य रूपमा काम गर्न आवश्यक पर्ने...
Source: Patient.info

✓ Found 1458 informative Q&A pairs
✓ Found 938 records from Patient.info

📋 DataFrame Shape: (3174, 10)
Average Question Length: 7.6 words
Average Answer Length: 39.9 words

7.4 Filtering and Searching

# Filter by domain
health_records = dataset.get_by_domain('स्वास्थ्य')
print(f"Health domain records: {len(health_records)}")

# Filter by question type
fact_based = [r for r in records 
              if 'तथ्यमा आधारित' in json.loads(r['metadata_json']).get('question_type', '')]
print(f"Fact-based questions: {len(fact_based)}")

# Search by keyword in question
def search_questions(dataset_obj, keyword):
    results = []
    for record in dataset_obj.records:
        conversations = record.get('conversations', [])
        for conv in conversations:
            if conv.get('from') == 'human' and keyword.lower() in conv.get('value', '').lower():
                results.append(record)
                break
    return results

cancer_questions = search_questions(dataset, 'क्यान्सर')
print(f"Records containing 'क्यान्सर' (cancer): {len(cancer_questions)}")

7.5 Data Export

# Export to CSV
df = dataset.to_dataframe()
df.to_csv('nepali_health_faq.csv', index=False, encoding='utf-8-sig')
print("✓ Exported to nepali_health_faq.csv")

# Export to JSON
with open('nepali_health_faq_processed.json', 'w', encoding='utf-8') as f:
    json.dump(records, f, ensure_ascii=False, indent=2)
print("✓ Exported to nepali_health_faq_processed.json")

# Export statistics
stats_df = pd.DataFrame([
    {'Metric': 'Total Records', 'Value': len(records)},
    {'Metric': 'Unique Sources', 'Value': len(set(s['source'] for s in records))},
    {'Metric': 'Avg Question Length', 'Value': f"{df['question_length'].mean():.1f} words"},
    {'Metric': 'Avg Answer Length', 'Value': f"{df['answer_length'].mean():.1f} words"}
])
stats_df.to_csv('dataset_statistics.csv', index=False)
print("✓ Exported statistics to dataset_statistics.csv")

7.6 Working with Conversations

def extract_all_qa_pairs(records):
    """
    Extract all Q&A pairs from records.
    """
    qa_pairs = []
    
    for record in records:
        conversations = record.get('conversations', [])
        
        for i, conv in enumerate(conversations):
            if conv.get('from') == 'human' and i + 1 < len(conversations):
                if conversations[i + 1].get('from') == 'gpt':
                    qa_pairs.append({
                        'record_id': record.get('id'),
                        'question': conv.get('value'),
                        'answer': conversations[i + 1].get('value'),
                        'source': record.get('source'),
                        'language': record.get('language')
                    })
    
    return qa_pairs

qa_pairs = extract_all_qa_pairs(records)
print(f"✓ Extracted {len(qa_pairs)} Q&A pairs")

# Create DataFrame of Q&A pairs
qa_df = pd.DataFrame(qa_pairs)
qa_df['question_word_count'] = qa_df['question'].apply(lambda x: len(x.split()))
qa_df['answer_word_count'] = qa_df['answer'].apply(lambda x: len(x.split()))

print(qa_df[['question', 'answer', 'source']].head())

8. DATASET CHARACTERISTICS AND QUALITY

Data Quality Features

Authentic Health Information: All content from real healthcare providers and medical organizations
Professional Sources: Mix of international (UK-based), regional (India), and local (Nepal) medical expertise
Diverse Medical Specialties: Covers 20+ different medical specialties and health topics
Consistent Formatting: Structured JSONL format with standardized metadata
Multilingual Metadata: Nepali questions and answers with detailed Nepali metadata
Comprehensive Conversation Structure: Question-answer pairs organized in conversation format

Language Characteristics

Script: Devanagari (देवनागरी) - Standard Nepali script
Language Code: npi (Nepali ISO 639-3 code)
Content Quality: Natural, fluent Nepali language with proper grammar and terminology
Medical Terminology: Uses standard medical terminology in Nepali

Data Distribution Insights

Heavily Health-Focused (95.7%): Ensures deep coverage of health topics
Balanced Behavior Types (45.9% + 42.9% = 88.8%): Informative and general information dominate
Question-Focused (89.2% fact-based): Practical, informative questions
Multi-Source (20+ sources): Reduces bias from single provider
Professional Expertise: From hospitals, clinics, international organizations, and government agencies

9. RECOMMENDED USE CASES

1. Training Conversational AI Systems

Build chatbots that answer health questions in Nepali
Train dialogue systems for patient support
Develop Q&A systems for medical advice

2. Information Retrieval Systems

Search engines for Nepali health information
Document retrieval systems for medical research
FAQ-based help systems

3. Natural Language Processing Tasks

Machine translation (English ↔ Nepali health content)
Text classification (medical topic categorization)
Named entity recognition (medical terms)
Question generation and understanding

4. Educational Applications

Medical education platforms in Nepali
Patient education materials
Health literacy programs

5. Healthcare Technology

Clinical decision support systems
Patient information databases
Telemedicine platforms serving Nepali-speaking populations

10. DATASET LIMITATIONS AND CONSIDERATIONS

Known Limitations

Language Scope: Nepali language only (no English or multilingual queries)
Question Type Dominance: 89.2% are fact-based, fewer procedural or advisory questions
Source Imbalance: RadFAQs (38.9%) and Patient.info (29.6%) dominate the dataset
Domain Concentration: 95.7% health-focused with limited diversity in other fields
Question Complexity: Most questions are relatively simple (average 7.6 words)

Recommendations for Use

For Training: Consider stratified sampling if using subset
For Evaluation: Test model performance across different sources
For Deployment: Verify appropriateness of UK-based health information for Nepali context
For Expansion: Supplement with Nepali-specific health information as needed

11. CITATION AND ATTRIBUTION

If you use this dataset in research or applications, please cite:

@dataset{nepali_health_faq_2024,
  title={Merged Nepali Health FAQ Dataset},
  author={Multiple Healthcare Organizations},
  year={2024},
  license={Apache-2.0},
  language={Nepali},
  records={3174},
  note={Compiled from 20+ healthcare providers including Patient.info, RadFAQs, Cancer Clinic Pune, and various Nepali health organizations}
}

Acknowledge these organizations when using the dataset:

Patient.info
RadFAQs (Radiology FAQs)
Cancer Clinic Pune
Chest Clinic of Kathmandu & Allergy Center
SAC Nepal
Fight Vitiligo
Mechi Eye Hospital
Nepal-based health organizations and government services

12. GETTING STARTED CHECKLIST

Download merged_nepali_health_faq_dataset.jsonl
Install Python 3.7+ and required libraries
Run dataset loader script to verify data integrity
Generate dataset statistics
Export to preferred format (CSV, JSON, DataFrame)
Analyze behavior and question type diversity
Filter data by domain/source for your use case
Prepare for model training or system development

13. TECHNICAL SPECIFICATIONS

File Format Details

Format: JSONL (JSON Lines) - one complete JSON object per line
Encoding: UTF-8 with Devanagari script support
File Size: ~50-100 MB (estimated)
Compression: Not compressed (recommend gzip for storage)

Required Fields in Each Record

{
  "id": "unique_identifier",
  "conversations": [
    {"from": "human", "value": "question_text"},
    {"from": "gpt", "value": "answer_text"}
  ],
  "source": "organization_name",
  "language": "ne",
  "language_code": "npi",
  "script": "Deva",
  "metadata_json": "{"behavior":"...", "question_type":"...", "domain":"..."}",
  "license": "Apache-2.0"
}

Optional Fields

source_name: Display name of source
source_split: train/test/validation split
condition: real/synthetic
generation_type: real/synthetic
url: Source URL (if available)

14. SUPPORT AND FEEDBACK

For questions, issues, or feedback:

Review the documentation above
Check your data loading code for errors
Verify file encoding is UTF-8
Ensure sufficient memory for large-scale processing
Test with small subsets first (e.g., 100 records)

15. QUICK REFERENCE TABLES

All Behaviors (5 types)

Behavior	Count	Use Case
Informative Answer	1,458	Detailed health information
General Information	1,363	Overview-level knowledge
Brief Answer	215	Quick facts
Direct Answer	100	Yes/no or one-liner
Short & Informative	38	Balanced response

All Question Types (6 types)

Type	Count	Example
Fact-Based	2,831	"क्यान्सर किन हुन्छ?" (Why does cancer occur?)
Interrogative	101	Questions seeking confirmation
Health-Specific	100	Disease-specific queries
Structured Q&A	100	Pre-formatted pairs
General	4	Broad health questions
Unclassified	38	No clear category

Top 10 Sources

Source	Records	Focus
RadFAQs	1,235	Medical imaging
Patient.info	938	General health
Cancer Clinic Pune	191	Oncology
Multiple (each <100)	810	Specialty topics

Dataset Version: v1.0
Last Updated: 2024
Format: JSONL (3,174 records)
Language: Nepali (नेपाली)
License: Apache-2.0 (Permissive)

Downloads last month: 23

Total file size:

7.98 MB