sabin1234/Nepali_SCRAP_Health_Instruction_Dataset

Nepali Health Instruction Dataset (REAL_SCRAP_HEALTH_COMPILE_DATASET)

A compiled instruction-following dataset of 87,182 Nepali-language question–answer pairs on public health. The records combine official Nepali health statistics, a US CDC county-level mortality table, an Economist / WHO vaccination scenario dataset, and model-generated and templated Q&A. Every record is a two-turn conversation (one human question, one assistant answer) stored as JSON Lines.

This README documents the file's structure, its composition, the diversity of behaviours and question patterns, known data-quality issues, and how to load it.

Read this first: Not all records are "real". The file name says REAL_SCRAP, but only 3.7% of records (3,245) are labelled condition = real. About 96% are synthetic or model-generated, and about 65% of the file is US county mortality data translated into Nepali. See Section 2.

---

---

1. At a Glance

Property	Value
File	REAL_SCRAP_HEALTH_COMPILE_DATASET.jsonl
File size	~150 MB
Records (lines)	87,182
Unique question–answer pairs	86,602 (580 exact duplicate copies)
Unique question strings	86,602
Turns per record	Exactly 2 (human → gpt)
Primary language	Nepali (ne, ISO code npi), Devanagari script (Deva)
Records where both turns contain Devanagari text	87,182 (100%)
Records tagged language = ne	86,595 (99.3%)
Records with missing metadata	587 (0.7%)
Number of distinct source datasets (source_name)	35, plus 587 rows with no source name
Licence tag	apache-2.0 on all rows (see Section 5)
Task type	instruction-following
Answer style	Short factual sentences, MCQ letters, numeric answers, and some long explanations (max 1,625 characters)

---

2. What Is Real and What Is Not

The condition and generation_type fields describe how each record was produced. The table below uses the values stored in the file.

condition	Records	Share	generation_type	What it means in this dataset
model-generated	57,648	66.1%	synthetic	Questions and answers written by an LLM from a source table. Covers the CDC block (gemini-3.5-flash-lite) and the health-awareness block (himalaya-gemma-4-e2b-it).
synthetic	25,702	29.5%	synthetic	Template-based or scenario-based questions built by the compiler from a source. Covers the Economist / WHO scenario block, the target-population MCQ block, the dengue block and the pharmacy exam block.
real	3,245	3.7%	real	Questions and answers derived from official Nepali government and survey tables (Department of Health Services, Ministry of Health and Population, Nepal DHS 2016, NHRC, and similar).
null	587	0.7%	null	No condition or source metadata recorded.

Key implications

• Only about 3.7% of records come from official Nepali tables. Treat the real label as the compiler's label. It describes the source data and has not been independently verified record by record.
• The model-generated and synthetic labels mean the question–answer text was produced by a program or a model, so it can contain generation errors. See Section 11.
• The largest block (CDC, 64.75% of records) is about United States county and state mortality statistics translated into Nepali. It is not Nepal data.

---

3. File Format and Field Reference

The file is JSON Lines: one JSON object per line, UTF-8 encoded, with CRLF (\r\n) line endings. Parsers should strip whitespace before calling json.loads.

3.1 Top-level fields

Field	Type	Always present?	Description
id	string	Yes	Record identifier. Not unique: 461 IDs appear exactly twice, each time for a different question. Do not use id as a primary key.
conversations	list of 2 objects	Yes	[{"from": "human", "value": ...}, {"from": "gpt", "value": ...}]. Always 2 turns.
source	string or null	Mostly	Human-readable source citation.
source_name	string or null	Mostly	Short internal name of the source block (e.g., target_population_fy2072_73).
source_repo	string or null	Mostly	Originating organisation or repository.
source_config	string or null	Mostly	Dataset configuration name (e.g., default, health_awareness, drug_poisoning_mortality).
source_split	string or null	Mostly	Split of the source dataset, usually train.
source_revision	string or null	Mostly	Source version tag (e.g., v1, v1.0, dohs_2015_16).
source_row_id	string or null	Mostly	Row identifier in the original source.
language	string or null	Mostly	ne (Nepali).
language_code	string or null	Mostly	npi (ISO 639-3 code for Nepali).
script	string or null	Mostly	Deva (Devanagari).
license	string	Yes	apache-2.0 on every row.
license_tier	string	Yes	permissive on every row.
task_type	string or null	Mostly	instruction-following.
generation_type	string or null	Mostly	real or synthetic.
condition	string or null	Mostly	real, synthetic or model-generated.
url	string	Yes	Always an empty string in this file.
metadata_json	string (JSON-encoded) or null	Mostly	A JSON string nested inside the record. Its keys depend on the source block (see 3.2).

3.2 Keys inside metadata_json

The metadata keys are not the same across the file. The table shows how many records carry each key.

Key	Records	Meaning
domain	57,648	Top-level domain: drug_poisoning_mortality (CDC block) or health_awareness.
model	57,648	Generator model: gemini-3.5-flash-lite or himalaya-gemma-4-e2b-it.
generation_number	57,648	Sequence number of the record within its generation run.
generation_domain	28,947	Nepali-language domain label (e.g., स्वास्थ्य = health).
generation_category	28,947	Category label (21 distinct values). See Section 9.
generation_sub_domain	28,947	Sub-category label (64 distinct values).
behavior	28,947	Behaviour label the response was meant to follow. See Section 6.
behavior_definition	28,947	Plain-language definition of the behaviour (in Nepali or English, depending on the block).
question_type	28,947	Question-type label (22 distinct values). See Section 7.
content_language	28,947	नेपाली (Nepali) or English. The English value is wrong for 256 records (see Section 11).
content_script	28,947	देवनागरी (Devanagari) or Latin. Same caveat as above.
english_content_allowed	28,947	Whether English content was permitted for the block.
question_length	25,602	Generation-time length constraint (text, not a measured value), e.g., "170 अक्षर" or "120 to 350 characters".
response_length	25,602	Generation-time length constraint for the answer.
maximum_response_sentences	25,602	1 (25,346 records) or 3 (256 records).
minimum_reasoning_dimensions	25,602	1 (24,597) or 2 (1,005).
minimum_complexity_score	25,602	1 (24,597) or 5 (1,005).
num_options	749	Number of options in multiple-choice questions (always 4).

3.3 Example record (trimmed)

{
  "id": "ari_pneumonia_0001",
  "conversations": [
    {"from": "human", "value": "प्रदेश १ मा तीव्र श्वासप्रश्वास संक्रमणको जोखिममा रहेका ५ वर्षमुनिका लक्षित जनसंख्या कति थियो?"},
    {"from": "gpt", "value": "प्रदेश १ मा तीव्र श्वासप्रश्वास संक्रमणको जोखिममा रहेका ५ वर्षमुनिका लक्षित जनसंख्या ४९४३०१ थियो।"}
  ],
  "source": "Disease Classification CB-IMNCI FCHV Services and Budget FY 2073/74",
  "source_name": "cb_imnci_fchv_budget",
  "source_repo": "Ministry of Health and Population",
  "language": "ne",
  "language_code": "npi",
  "script": "Deva",
  "license": "apache-2.0",
  "task_type": "instruction-following",
  "generation_type": "real",
  "condition": "real",
  "url": "",
  "metadata_json": "{\"generation_domain\": \"स्वास्थ्य\", \"behavior\": \"तथ्यमा आधारित उत्तर दिने\", \"question_type\": \"तथ्यमा आधारित प्रश्न\", ...}"
}

English gloss of the example: "How large was the target population of children under 5 at high risk of acute respiratory infection (ARI) in Province 1?" → "In Province 1, the target population of children under 5 at high risk of ARI was 494,301."

---

4. Composition by Source

The table lists every source block in the file, sorted by size. Question and answer lengths are medians in Unicode characters.

Source (source_name)	Records	Share	Condition	Generator model	Domain (metadata.domain or generation_domain)	Question median	Answer median
Centers for Disease Control and Prevention	56,448	64.75%	model-generated	gemini-3.5-flash-lite	drug_poisoning_mortality	132	141
The Economist / WHO Vaccination Data	24,597	28.21%	synthetic	—	स्वास्थ्य (health)	172	22
nepali_health_awareness_sft	1,200	1.38%	model-generated	himalaya-gemma-4-e2b-it	health_awareness	53	986.5
communicable_outbreaks_2073_74	986	1.13%	real	—	स्वास्थ्य (health)	61	43
target_population_fy2072_73	749	0.86%	synthetic	—	स्वास्थ्य जनसांख्यिकी (health demographics)	171	8
(missing)	587	0.67%	—	—	—	67	66
fchv_district_2072_73	402	0.46%	real	—	स्वास्थ्य (health)	51	43
cb_imnci_fchv_budget	300	0.34%	real	—	स्वास्थ्य (health)	78	78.5
dengue_positive_cases_fy2071_74	256	0.29%	synthetic	—	Public Health	93	86.5
shsp_enrollment	183	0.21%	real	—	स्वास्थ्य (health)	84	38
women_bank_mobile	176	0.20%	real	—	स्वास्थ्य तथा सामाजिक (health and social)	69.5	69
nhrc_research	107	0.12%	real	—	स्वास्थ्य (health)	82	61
pharmacy_licensing_exam_faq	100	0.11%	synthetic	—	सार्वजनिक स्वास्थ्य परीक्षा तथ्याङ्क (public health exam data)	45.5	34
dohs_province_health	93	0.11%	real	—	स्वास्थ्य (health)	67	64
hospital_cost	92	0.11%	real	—	स्वास्थ्य (health)	51.5	56.5
fchv_home_delivery	89	0.10%	real	—	स्वास्थ्य (health)	79	41
dohs_health_stats	80	0.09%	real	—	स्वास्थ्य (health)	84	83
dog_bite_rabies	74	0.08%	real	—	स्वास्थ्य (health)	55	53
hospital_employees	71	0.08%	real	—	स्वास्थ्य (health)	54	52
npafp	68	0.08%	real	—	स्वास्थ्य (health)	75	41
malaria_trend	68	0.08%	real	—	स्वास्थ्य (health)	56	56.5
snake_bite	55	0.06%	real	—	स्वास्थ्य (health)	54	54
diarrhoea_treatment	54	0.06%	real	—	स्वास्थ्य (health)	92	93
immunization	49	0.06%	real	—	स्वास्थ्य (health)	60	62
tb_notification	39	0.04%	real	—	स्वास्थ्य (health)	79	78
measles_rubella	34	0.04%	real	—	स्वास्थ्य (health)	34	33
diarrhoea_incidence	34	0.04%	real	—	स्वास्थ्य (health)	97.5	98
fp_modern	33	0.04%	real	—	स्वास्थ्य (health)	81	84
health_services_trend	29	0.03%	real	—	स्वास्थ्य (health)	50	54
ayurvedic_council	28	0.03%	real	—	स्वास्थ्य (health)	91.5	75.5
adolescent_abortion	26	0.03%	real	—	स्वास्थ्य (health)	65	73
adolescent_anc	25	0.03%	real	—	स्वास्थ्य (health)	95	96
phc_orc	18	0.02%	real	—	स्वास्थ्य (health)	83.5	87
pmtct_trend	16	0.02%	real	—	स्वास्थ्य (health)	64	63.5
patient_types	10	0.01%	real	—	स्वास्थ्य (health)	52.5	52.5
nurses_reg	6	0.01%	real	—	स्वास्थ्य (health)	47.5	49.5
Total	87,182	100%

Observations

• The top two sources make up 92.96% of the file. The CDC block and the Economist / WHO block together are 80,000+ records built from a small number of templates.
• The long tail of 29 small real sources adds Nepali-specific topics (FCHV volunteers, ARI, malaria, TB, snake bite, rabies, abortion and adolescent health, workforce registration) but only 3.7% of the file in total.
• The nepali_health_awareness_sft block is the only source with long answers (median 986.5 characters), so it is the main source of multi-sentence explanations.

---

5. Language, Script and Licensing

Attribute	Distribution	Notes
language / language_code	ne / npi for 86,595 records	587 records have no language tag.
script	Deva for 86,595 records	Same 587 records have no script tag.
Question text contains Devanagari	87,182 of 87,182 (100%)	Every question is in Devanagari.
Answer text contains Devanagari	87,182 of 87,182 (100%)	Some answers still contain English words or uppercase romanised names (see below).
Answers containing Latin letters	105 records total	54 in nepali_health_awareness_sft, 43 in target_population_fy2072_73, 8 in pharmacy_licensing_exam_faq.
Digit forms	Mixed	Nepali digits (०–९) are the norm, and some answers use ASCII digits. Use Unicode-aware normalisation before any numeric parsing.
license	apache-2.0 on all 87,182 rows	This is a dataset-level tag applied by the compiler. Check the upstream terms of each source (see below).

Licensing note. The license field says apache-2.0 for every row. The underlying sources are published by different bodies:

• The CDC data is a US federal government publication.
• The Economist / WHO data is published by the Economist Intelligence Unit and the WHO.
• The Nepali sources are Government of Nepal publications (Department of Health Services, Ministry of Health and Population, Central Bureau of Statistics, NHRC).

Before any commercial or redistribution use, check each upstream publisher's terms. The apache-2.0 tag does not override them.

---

6. Behaviour Diversity

The behavior metadata field records the response style each block was designed to teach. Five behaviour labels exist, and 58,235 records (66.8%) have no behaviour label at all.

behavior (original)	English gloss	Records	Share	Source blocks	Behaviour definition (English gloss)
बहुविकल्पीय प्रश्नको उत्तर दिने	Answer multiple-choice questions	24,597	28.2%	Economist / WHO	Choose the one correct option from four that matches the data.
तथ्यमा आधारित उत्तर दिने	Give fact-grounded answers	3,245	3.7%	All real blocks	Keep the core fact intact and answer in clear, standard Nepali.
select correct option	Select the correct option	749	0.9%	target_population_fy2072_73	Choose the single correct option from four and state it clearly.
short factual answer	Short factual answer	256	0.3%	dengue_positive_cases_fy2071_74	Give a concise, data-grounded answer that keeps the key fact.
प्रत्यक्ष र प्राकृतिक उत्तर दिने	Give direct, natural answers	100	0.1%	pharmacy_licensing_exam_faq	Answer directly and naturally, only from the values given, in plain Nepali.
(no label)	—	58,235	66.8%	CDC, health awareness, missing-metadata rows	Not specified.

What the behaviours teach, in practice

• Exact-option selection (25,346 records) teaches the model to pick one of four labelled options and, in most cases, to return the letter with the value.
• Fact-grounded short answers (3,245 real records, plus dengue and pharmacy) teach brief, data-anchored replies in Nepali.
• Free-form explanation (the health-awareness block, 1,200 records) teaches multi-sentence, general-audience health education.
• Scenario reasoning (the Economist / WHO block) asks for numerical outcomes under hypothetical coverage levels, so it is the closest the file gets to reasoning over a model.

Answer-behaviour summary

Answer behaviour	Count	Example (English gloss)
Other free-text sentence answer	58,442	"In Province 2, the figure was 613,361."
MCQ letter only	24,597	सही उत्तर: ग) ० । → "Correct answer: c) 0."
Short numeric answer	2,194	"1, 0 and 0 people died."
Letter plus option text	749	घ. ६६२९ → "d. 6629"
Long explanatory paragraph (≥300 characters)	1,200	General health-awareness explanation.

---

7. Question-Pattern Diversity

7.1 Question types (metadata question_type)

The metadata defines 22 question-type labels. The 12 most common are shown below. The remaining 10 labels account for the other 197 labelled records.

question_type	Records	Source	Description
cohort_size	3,509	Economist / WHO	Asks for the size of a target cohort under a scenario.
cancer_prevented	3,493	Economist / WHO	Asks how many cancer cases were prevented.
deaths_prevented	3,485	Economist / WHO	Asks how many deaths were prevented.
possible_cancer_cases	3,484	Economist / WHO	Asks for possible cancer cases.
possible_cancer_deaths	3,473	Economist / WHO	Asks for possible cancer deaths.
prevented_case_percentage	3,295	Economist / WHO	Asks for the percentage of cases prevented.
तथ्यमा आधारित प्रश्न (fact-based question)	3,245	real blocks	Direct question about a fact in an official table.
coverage	3,150	Economist / WHO	Asks about vaccination coverage.
multiple-choice grounded question	749	target_population_fy2072_73	Four-option question grounded in a population table.
cancer_vs_deaths_prevented	511	Economist / WHO	Compares cancer and death outcomes.
complex grounded open-ended question	256	Dengue	Open-ended question that needs a short data-grounded answer.
प्रत्यक्ष संवादात्मक प्रश्न (direct conversational question)	100	Pharmacy exam FAQ	Conversational question in exam-prep style.

7.2 Surface patterns in the question text

These counts are computed by keyword matching on the question text. They are approximate. The keyword lists are given so you can reproduce the counts.

Pattern	Keywords used	Records	Share	Notes
Contains a year or fiscal-year reference	वि.सं., आर्थिक वर्ष, any four-digit year	83,370	95.6%	Mostly Economist (वि.सं. 2021–2030) and FY 2071/72–2073/74 references.
Asks for a count ("how many")	कति	71,316	81.8%	The dominant question form.
Scenario or counterfactual	परिदृश्य (scenario), यदि (if)	24,618	28.2%	Almost all from the Economist / WHO block.
Multiple-choice with four options	Source-level count, not keyword-based	25,346	29.1%	24,597 Economist (letter answer) and 749 target-population (letter + text).
Comparison between two or more entities	मध्ये (between/among), बीच (between), तुलना (comparison), भिन्नता (difference), भन्दा (than)	5,112	5.9%	Examples: which of two districts has more X.
Asks about change over time	परिवर्तन (change), बढ (increase), घट (decrease)	1,737	2.0%	Dengue and trend-based questions.
Explanatory or "how/why"	किन, कसरी (how), के हो (what is), व्याख्या (explain)	1,479	1.7%	Mostly the health-awareness block.
Asks which entity is higher or lower	कुन ... बढी/कम (which ... more/less)	512	0.6%	Small share; these ask for a direct comparison of two or more entities.

7.3 Question-length and structure

Question length (Unicode characters)	Records	Share
Under 50	1,047	1.2%
50–99	4,178	4.8%
100–149	56,875	65.2%
150–199	24,151	27.7%
200 or more	931	1.1%

Most questions are 100–199 characters. That is a single well-formed sentence, often with a year, an entity and a metric. The file has very little variation in question length, and only about 1% of questions exceed 200 characters.

7.4 Template structure

The two largest blocks (CDC and Economist / WHO, together 92.96% of records) use a small number of sentence templates, with the entity, year and metric swapped in. Typical openings:

Opening (first two words)	Records	Block
वि.सं. २०२X (Vikram Samvat year, 2021–2030)	22,741 (sum of the ten most common openings)	Economist / WHO
आर्थिक वर्ष (fiscal year)	2,117	Nepali official tables
टेक्सास राज्यको (of the state of Texas)	941	CDC
उत्तरी क्यारोलिना (North Carolina)	583	CDC
जर्जिया राज्यको (of the state of Georgia)	578	CDC
भर्जिनिया राज्यको (of the state of Virginia)	500	CDC
२००४ मा (in 2004)	457	CDC

What this means for diversity: lexical and structural variety is low inside the two largest blocks. Entities, years and numbers vary a lot, but the sentence frames repeat. If you train on the full file, the model will see the same frames thousands of times.

7.5 Multiple-choice format (4 options each)

Format	Records	Example answer
Letter in a "correct answer" sentence	24,597	सही उत्तर: घ) ५००.०६ । (Correct answer: d) 500.06)
Letter followed by the option text	749	ख. सुनसरी (b. Sunsari)

The four options appear to be drawn from the same kind of value as the answer (for example, numbers of the same kind or district names), which makes them more useful as distractors than random values. Check a sample before relying on this for evaluation.

---

8. Answer Diversity

8.1 Answer format

Format	Records	Share
Sentence answer (free text, 1–2 sentences)	58,442	67.0%
सही उत्तर: <letter>) (MCQ correct-answer sentence)	24,597	28.2%
Short numeric answer (under 60 characters with digits)	2,194	2.5%
Long explanatory paragraph (≥300 characters)	1,200	1.4%
Letter plus option text	749	0.9%

8.2 Answer length (Unicode characters)

Answer length	Records	Share
Under 25	23,376	26.8%
25–74	5,019	5.8%
75–149	43,832	50.3%
150–499	13,757	15.8%
500 or more	1,198	1.4%

The longest answer is 1,625 characters. Long answers appear almost only in the health-awareness block.

8.3 Answer style

• Restating style. Most sentence answers restate the entity and the metric from the question, then give the number. This makes answers easy to evaluate against the question.
• Bare-number style. About a quarter of answers are only a number or a short phrase. Use these for exact-match evaluation, but not as examples of good conversational style.
• Explanatory style. The health-awareness answers are general-audience explanations (on breathing, adolescent health, and similar topics). They are the only long-form, instruction-like answers in the file.
• Scenario-style numbers. The Economist / WHO answers are numbers under a hypothetical scenario, usually given with the option letter.

---

9. Topic and Domain Diversity

9.1 Top-level domain (metadata generation_domain)

Domain (Nepali)	English gloss	Records
स्वास्थ्य	Health	27,666
स्वास्थ्य जनसांख्यिकी	Health demographics	749
Public Health	Public health	256
स्वास्थ्य तथा सामाजिक	Health and social	176
सार्वजनिक स्वास्थ्य परीक्षा तथ्याङ्क	Public health exam data	100

9.2 Category (metadata generation_category, top 12 of 21)

Category (Nepali)	English gloss	Records
एचपीभी तथा गर्भाशयको मुखको क्यान्सर	HPV and cervical cancer	24,597
संक्रामक रोग प्रकोप	Communicable disease outbreaks	986
महिला सामुदायिक स्वास्थ्य स्वयंसेविका	Female Community Health Volunteers (FCHV)	402
स्वास्थ्य सेवा तथ्यांक	Health service statistics	332
सीबी–आईएमएनसीआई तथा एफसीएचभी सेवा	CB-IMNCI and FCHV services	300
जिल्ला सूचकहरू	District indicators	278
स्वास्थ्य बिमा तथा प्रसूति सहयोग	Health insurance and maternity support	272
क्षयरोग तथा स्वास्थ्य सेवा तथ्यांक	Tuberculosis and health service statistics	271
Dengue Surveillance	Dengue surveillance	256
आयुर्वेद तथा स्वास्थ्य सेवा	Ayurveda and health services	188
महिलाको बैंक खाता तथा मोबाइल प्रयोग	Women's bank accounts and mobile use	176
मातृ स्वास्थ्य तथा रोग नियन्त्रण	Maternal health and disease control	145

9.3 Health topics covered by the real and small blocks

The sources cover a broad set of Nepali public-health topics. Examples by source:

Topic	Source blocks (examples)
Community health volunteers (FCHV)	fchv_district_2072_73, fchv_home_delivery
Child pneumonia and IMNCI (ARI)	cb_imnci_fchv_budget
Outbreaks (diarrhoea, influenza, scrub typhus, leptospirosis, food poisoning)	communicable_outbreaks_2073_74, diarrhoea_incidence, diarrhoea_treatment
Dengue	dengue_positive_cases_fy2071_74
Malaria	malaria_trend
Tuberculosis	tb_notification
Immunisation and measles / rubella	immunization, measles_rubella
Family planning	fp_modern
Maternal and adolescent health	adolescent_anc, adolescent_abortion, pmtct_trend
Rabies, snake bite	dog_bite_rabies, snake_bite
Health insurance (social health security)	shsp_enrollment
Hospital cost and staffing	hospital_cost, hospital_employees, patient_types, nurses_reg
Women's banking and mobile use	women_bank_mobile
Ayurveda and workforce registration	ayurvedic_council
Pharmacy licensing exam	pharmacy_licensing_exam_faq
Health research	nhrc_research
Province-level health statistics	dohs_province_health, dohs_health_stats
Health awareness (general education)	nepali_health_awareness_sft
HPV vaccination and cervical cancer (scenario)	The Economist / WHO Vaccination Data
Drug poisoning mortality, US counties (translated)	Centers for Disease Control and Prevention

Breadth vs depth. The real blocks are broad in topic, but each block is small (typically 10–400 records). The HPV / cervical-cancer and US mortality blocks are deep but narrow.

9.4 Geographic and population entities

• Nepal. Questions name provinces (e.g., Province 1, Province 2) and districts across the country, including Ilam, Morang, Sunsari, Kaski, Lamjung, Rupandehi, Dhankuta, Kailali, Dadeldhura, Jajarkot, Humla, Dolpa, Jumla, Dang, Doti, Sindhupalchok, Nuwakot, Achham, Palpa, Pyuthan, Panchthar, Myagdi, Mustang, Lalitpur and Kapilvastu.
• United States. The CDC block names many US states and counties (e.g., Texas, Georgia, North Carolina, Virginia, Kentucky, Morgan County and Marion County, Indiana), written in Nepali.
• Countries in the Economist / WHO block. Scenario questions name countries such as Lesotho and Sweden.

---

10. Temporal and Geographic Coverage

10.1 Time references

Reference type	Values in the file	Notes
Fiscal year (Nepali official)	9 distinct values, including २०७३/७४ (1,056 mentions), २०७२/७३ (915), २०७१/७२ (150)	Nepali fiscal years run mid-July to mid-July.
Vikram Samvat year (वि.सं.)	2021–2030	Used in the Economist / WHO block. See the caveat below.
Gregorian year (ASCII or Devanagari digits)	Mostly CDC years, e.g., 2003, 2004, 2005, 2007, 2008, 2014	The CDC block uses Gregorian years.

Calendar caveat. The Economist / WHO block writes its years as वि.सं. (Vikram Samvat). Those values look like projection years from a Gregorian WHO / Economist scenario (2021 to 2030). If that is right, the label is a calendar error. Confirm the calendar before you use these years for any date arithmetic.

10.2 Geographic scope

Scope	Records (approx.)	Share (approx.)
United States (county and state level, translated)	56,448	64.75%
Global / countries (HPV scenario, Economist / WHO)	24,597	28.21%
Nepal (national, provincial and district)	About 4,250	About 4.9%
Unspecified or general health awareness	1,200 health-awareness + 587 missing-metadata rows	About 2.1%

The Nepal figure is the sum of the real, target-population and dengue blocks (3,245 + 749 + 256 = 4,250). It is an approximation, because a few real questions may not name a location.

---

11. Data Quality and Known Issues

The following issues were found in the file. Counts are exact unless marked otherwise.

#	Issue	Records affected	Impact	Recommended handling
1	Duplicate id values. The same id is used for two different questions.	461 IDs (922 records)	Joins, deduplication and lookups by id will be wrong.	Deduplicate by a hash of (question, answer). Use a new row index as the key.
2	Exact duplicate question–answer pairs.	580 extra copies	Inflates counts and can leak between train and test splits.	Drop duplicates before splitting.
3	Missing metadata. source, source_name, condition, task_type, metadata_json are null.	587 (0.67%)	These rows cannot be filtered by source or condition.	Keep them if you only need Q&A text. Exclude them if you need provenance.
4	Wrong language label. Dengue rows are labelled content_language = English and content_script = Latin, but the text is Nepali in Devanagari.	256	Filtering by these metadata values gives the wrong subset.	Do not filter on content_language / content_script for these rows.
5	Answers in English instead of Nepali in the target-population MCQ block. The letter is correct, but the district name is in uppercase English (e.g., ख. SUNSARI for option ख. सुनसरी).	42	Mixed-language answers.	Normalise to the Nepali option text, or exclude these rows if you need pure Nepali answers.
6	Spelling variants or garbled option text in the target-population MCQ block. The answer letter is correct, but the answer text differs from the option text (e.g., सिराहा vs सिरहा; option कम <> with stray markup).	39	Exact-match evaluation will mark correct answers wrong.	Match on option letter instead of text.
7	Stray markup (<>, </>) inside an option.	1	Looks broken if rendered.	Strip the markup.
8	Garbled or ungrammatical generated text in the health-awareness block. Example: the word किशोरकिशोवृद्धावस्था mixes "adolescence" and "old age" in one word.	Not measured across the block; observed in spot checks	Teaches incorrect Nepali.	Review or filter the block before training.
9	Answers that may not match the hypothetical in the question. Example from the pharmacy block: a "what if five more candidates pass" question is answered with a pass rate (101 of 110 = 91.82%) without clearly showing the changed numbers.	Not measured across the block; observed in spot checks	Teaches terse or off-target answers.	Verify numbers against the question before training.
10	Template repetition. The CDC and Economist / WHO blocks use a small set of sentence frames.	80,000+ records	Low structural diversity; the model may over-learn the frames.	Subsample by source, or add paraphrased variants.
11	Empty url field.	All 87,182	No source links are available from the file.	Use the source and source_repo fields for citation instead.
12	Calendar labelling in the Economist / WHO block (वि.सं. used for likely Gregorian years).	24,597	Date arithmetic will be wrong if the calendar is misread.	See Section 10.1.

What was checked and what was not

• Checked: record structure, field types, JSON validity (all 87,182 lines parse), duplicate IDs and duplicate QA pairs, language and script tags, the answer consistency of all 24,597 Economist MCQ items (100% consistent with their options), and the answer consistency of all 749 target-population MCQ items (letters all valid; 668 exact text matches; 42 English romanised answers; 39 spelling or markup mismatches).
• Not checked: the factual correctness of the answers against the original source tables, for all records. The spot checks described above found errors, so do not assume the numbers are correct without verification.

---

12. Loading the Dataset

The file is about 150 MB. You can load it in three ways, depending on what you need. The demo script load_dataset_demo.py in this folder contains all of the code below and was tested against the file.

12.1 Requirements

# Standard library only for the core loader.
# Optional, for the pandas and Hugging Face examples:
pip install pandas datasets

12.2 Core loader (standard library, memory-friendly)

import json
from collections import Counter
from pathlib import Path

DATA_PATH = Path("REAL_SCRAP_HEALTH_COMPILE_DATASET.jsonl")


# 1. Stream records one at a time (recommended for a 150 MB file)
def stream_records(path=DATA_PATH):
    with open(path, encoding="utf-8") as f:
        for line in f:
            line = line.strip()          # the file uses CRLF line endings
            if line:
                yield json.loads(line)


# 2. Load everything into memory (fine on most laptops for this size)
def load_all(path=DATA_PATH):
    return list(stream_records(path))


# 3. metadata_json is a JSON *string* nested inside the record. Decode it.
def parse_metadata(record):
    raw = record.get("metadata_json")
    if raw is None:
        return {}
    if isinstance(raw, dict):
        return raw
    try:
        return json.loads(raw)
    except json.JSONDecodeError:
        return {}


# 4. Flatten one record into a row with the fields you usually need
def flatten(record):
    convs = record["conversations"]
    human = next(c["value"] for c in convs if c["from"] == "human")
    gpt = next(c["value"] for c in convs if c["from"] == "gpt")
    meta = parse_metadata(record)
    return {
        "id": record["id"],
        "question": human,
        "answer": gpt,
        "source_name": record.get("source_name"),
        "condition": record.get("condition"),
        "generation_type": record.get("generation_type"),
        "task_type": record.get("task_type"),
        "language_code": record.get("language_code"),
        "behavior": meta.get("behavior"),
        "question_type": meta.get("question_type"),
        "domain": meta.get("domain") or meta.get("generation_domain"),
        "model": meta.get("model"),
    }


# 5. Filter by any flattened field
def filter_records(records, **criteria):
    return [r for r in records if all(flatten(r).get(k) == v for k, v in criteria.items())]


# 6. Convert to chat-style messages (for fine-tuning frameworks)
def to_messages(record, system_prompt=None):
    role_map = {"human": "user", "gpt": "assistant"}
    msgs = []
    if system_prompt:
        msgs.append({"role": "system", "content": system_prompt})
    for turn in record["conversations"]:
        msgs.append({"role": role_map[turn["from"]], "content": turn["value"]})
    return {"id": record["id"], "messages": msgs}


if __name__ == "__main__":
    # Count by condition, streaming (no full load)
    counts = Counter(rec.get("condition") for rec in stream_records())
    print("By condition:", dict(counts))

    # Load one source and show a sample
    records = load_all()
    mcq = filter_records(records, source_name="target_population_fy2072_73")
    print("Target-population MCQ records:", len(mcq))
    print(flatten(mcq[0]))
    print(to_messages(mcq[0]))

Expected output (abridged)

By condition: {'real': 3245, 'synthetic': 25702, 'model-generated': 57648, None: 587}
Target-population MCQ records: 749

12.3 pandas

import json
import pandas as pd

rows = []
with open("REAL_SCRAP_HEALTH_COMPILE_DATASET.jsonl", encoding="utf-8") as f:
    for line in f:
        if not line.strip():
            continue
        r = json.loads(line)
        rows.append({
            "id": r["id"],
            "question": r["conversations"][0]["value"],
            "answer": r["conversations"][1]["value"],
            "source_name": r.get("source_name"),
            "condition": r.get("condition"),
        })

df = pd.DataFrame(rows)
print(df.shape)                           # (87182, 5)
print(df["condition"].value_counts(dropna=False))
print(df["source_name"].value_counts(dropna=False).head(5))

Expected output (abridged)

(87182, 5)
condition
model-generated    57648
synthetic          25702
real                3245
NaN                  587

12.4 Hugging Face datasets

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files="REAL_SCRAP_HEALTH_COMPILE_DATASET.jsonl",
    split="train",
)

print(ds)                  # Dataset({features: [...19 columns], num_rows: 87182})
print(ds[0]["conversations"])

# Filter by source
mcq = ds.filter(lambda r: r["source_name"] == "target_population_fy2072_73")
print(len(mcq))            # 749

Note: datasets leaves metadata_json as a string. Decode it with json.loads when you need the fields inside it.

12.5 Quick statistics

import json
from collections import Counter

with open("REAL_SCRAP_HEALTH_COMPILE_DATASET.jsonl", encoding="utf-8") as f:
    rows = [json.loads(l) for l in f if l.strip()]

print("Records:", len(rows))
print("Unique QA pairs:", len({(r["conversations"][0]["value"], r["conversations"][1]["value"]) for r in rows}))
print("Duplicate IDs:", sum(1 for v in Counter(r["id"] for r in rows).values() if v > 1))
print("Top sources:", Counter(str(r.get("source_name")) for r in rows).most_common(3))

12.6 Converting to a training format

# Alpaca-style instruction format
def to_alpaca(record):
    return {
        "instruction": record["conversations"][0]["value"],
        "input": "",
        "output": record["conversations"][1]["value"],
    }

# Chat-style format is shown in section 12.2 (to_messages).

---

13. Recommended Filters and Recipes

Goal	Filter or action
Only human-derived or official-table data	condition == "real" (3,245 records)
Pure Nepali answers without English	Exclude the 42 English-answer MCQ rows (target_population_fy2072_73 with ASCII answers)
Multiple-choice evaluation	question_type == "multiple-choice grounded question" (749), or the Economist block (24,597)
Open-ended explanation	source_name == "nepali_health_awareness_sft" (1,200, long answers)
Short factual QA	behavior == "तथ्यमा आधारित उत्तर दिने" (3,245 real records)
Exclude the largest translated block	source_name != "Centers for Disease Control and Prevention" (drops 56,448 rows)
Deduplicate	Hash (question, answer) and keep one copy per hash
Keep Nepal-only data	source_name in the Nepal source list (see Section 4)
Remove missing-metadata rows	source_name is not None (drops 587)

Balancing. The file is heavily imbalanced: two blocks make up 93% of records. For training, subsample the large blocks or weight the small real blocks more heavily. Otherwise the model will mostly learn the CDC and Economist frames.

---

14. Intended Use and Limitations

Intended use

• Research and experimentation on Nepali instruction tuning and evaluation.
• Building Nepali health question-answering benchmarks, after you have verified the answers you intend to use.
• Studying how templated and model-generated data affects a model's language and reasoning behaviour.

Limitations

• Not a clinical or medical reference. Numbers and statements are not validated against source documents. Do not use the dataset for diagnosis, treatment or policy decisions without checking the sources.
• Mostly synthetic. About 96% of records are model-generated or template-generated. Answers can be wrong, incoherent or ungrammatical (see Section 11).
• Not representative of Nepal. About two-thirds of the file is US data, and the remaining Nepal material is a small set of topics.
• Limited question diversity. Two blocks use a handful of sentence templates. The question length is narrow (mostly 100–199 characters).
• Language mixing. Some answers contain English words or romanised names.
• Calendar ambiguity. वि.सं. years in the Economist / WHO block may be Gregorian. See Section 10.1.
• Licensing. The apache-2.0 tag is applied at the dataset level. Check the upstream publisher terms before redistribution or commercial use.

---

15. How These Statistics Were Computed

Every number in this README was computed directly from the JSONL file with Python's standard library (json, re, collections, statistics). The checks were:

• Parsing all 87,182 lines, which all parsed successfully.
• Counting the values of each top-level and metadata_json field.
• Cross-tabulating source_name, condition, generation_type, model and domain.
• Counting exact duplicate (question, answer) pairs and duplicate id values.
• Measuring character lengths of questions and answers (Unicode code points).
• Checking the MCQ answers in the Economist and target-population blocks against the option text in each question.
• Keyword-based pattern counts (Section 7.2). These are approximate, and the keyword lists are given so you can reproduce them.

Statistics marked "about" or "approximately" are estimates. All other counts are exact.

---

Files in this folder

File	Purpose
README.md	This document.
load_dataset_demo.py	Tested loader script with streaming, flattening, filtering and chat-format conversion.

The dataset itself (REAL_SCRAP_HEALTH_COMPILE_DATASET.jsonl) is not copied into this folder. Place it next to the script before you run it.

Drug Overdose and HIV / Mental Health Q&A Datasets (Nepali)

Two Nepali question-answer datasets built from public-health awareness material: a drug overdose pamphlet and an HIV and mental health poster. Each file has 800 single-turn records in ShareGPT format.

This is health content. The answers are not medical advice and have not been checked by a clinician. Section 2 lists the specific safety points. Read Section 6 before using these files in any product, especially the HIV / mental health file, which has a strong yes-bias.

---

1. Contents

File	Rows	Size	Material it was built from	License field
drugs_overdose_nepali_qa_.jsonl	800	1.27 MB	Drug overdose awareness pamphlet	CC0: Public Domain (public health pamphlet)
hiv_mental_health_nepali_qa_.jsonl	800	1.21 MB	HIV and mental health poster	Apache-2.0

Both files are Nepali (language_code = npi, Devanagari). The text includes some English terms and acronyms (Naloxone, PWID, MSM/TG, and so on).

1.1 Source and license checks

Item	Drugs file	HIV file
source	ncasc-nepal/drugs-overdose-pamphlet	ncasc-nepal/hiv-and-mental-health-poster
source_repo	Nepal Government / Health related pamphlet	National Centre for AIDS and STD Control, Nepal
license	CC0: Public Domain	Apache-2.0
condition / generation_type	real / real	real / real

Three things to check before use:

1. Attribution is inconsistent in the drugs file. The source_repo says Nepal Government, while the source path points to NCASC (National Centre for AIDS and STD Control). Confirm which body published the pamphlet.
2. The license may not fit the source. A government or NCASC pamphlet is not usually released under Apache-2.0 or CC0. Check the publisher's copyright and reuse terms before redistribution.
3. These are not dated. The records do not say which edition of the pamphlet or poster was used. Hotline numbers and treatment information change over time.

---

2. Safety review of the content

These are the points checked in the answers. The review was automated plus sampling, not a full clinical review.

2.1 Drugs overdose file

Check	Result
Specific doses, quantities, or units (mg, ml, tablets)	None. No answer contains a number.
Instructions for obtaining or preparing drugs	None found.
Self-harm or suicide content	None found.
Recognition of overdose signs	Present. Examples: difficulty breathing, unconsciousness, blue lips or nails, sweating, vomiting, abnormal heartbeat.
First-aid guidance	Present. Call an ambulance, place the person on their side, check breathing, use naloxone if available, do not leave the person alone.
Harm reduction	Present. Avoid taking a large amount, avoid mixing drugs, avoid using alone, use sterile syringes, know about naloxone.

The file is consistent with public-health overdose guidance. It contains no dosing or method information, so it is suitable for awareness and first-aid content. It does not name a hotline or a clinical contact.

2.2 HIV and mental health file

Check	Result
Crisis line	Suicide prevention helpline 1166 appears in 221 answers. This is the only number in the file. Verify it against the current official source before use.
Method or means content	The keyword scan and sampled answers found no means or method descriptions.
Stigmatising words (for example, "mad")	None found.
Key-population acronyms	PWID (82), MSM (67), TG (67), MSM/TG (40), MLM (33). These are used without expansion.

The file has no dosing content and includes a crisis line, which are positives. The problems in Section 6 are about accuracy and how the answers are framed.

---

3. Record schema

Both files use the same structure. Neither has the fact_key, related_fund, or question_type fields found in the other datasets in this collection.

Field	Value / meaning
id, source_row_id	Drugs: drugs_overdose_...; HIV: hiv_mental_health_.... Unique per file.
conversations	Two turns: human question, then GPT answer.
source, url	As in Section 1.1.
source_name	drugs_overdose_pamphlet_nepal / hiv_and_mental_health_poster_nepal
language, language_code, script	ne, npi, Deva
condition, generation_type	real, real
task_type	instruction-following
license, license_tier	See Section 1.1.
metadata_json	JSON string. Parse with json.loads.

3.1 Fields inside metadata_json

Field	Drugs file	HIV file
generation_domain	Public Health / Drug Overdose Awareness	Public Health / HIV and Mental Health
generation_category	Drugs Overdose Pamphlet	HIV and Mental Health Poster
behavior	Provide factual answer derived only from the pamphlet	Provide factual answer derived only from the poster
content_language, content_script	नेपाली, देवनागरी	नेपाली, देवनागरी
english_content_allowed	False	False

Every metadata value is the same across the records in each file. The behaviour is stated as "derived only from the pamphlet" or "derived only from the poster", but the records do not say which page or section each answer came from.

3.2 Example record (drugs file)

{
  "conversations": [
    {"from": "human", "value": "जानकारी दिनुहोस्: ड्रग ओभरडोज भएमा के गर्नुपर्छ?"},
    {"from": "gpt", "value": "तत्काल एम्बुलेन्स बोलाउनुहोस्, व्यक्तिलाई साइडमा सुताउनुहोस्, सासको जाँच गर्नुहोस्, Naloxone भए प्रयोग गर्नुहोस् र कहिल्यै एक्लै नछोड्नुहोस्। यो आधिकारिक जानकारी हो।"}
  ],
  "metadata_json": "{\"behavior\": \"Provide factual answer derived only from the pamphlet\", \"...\": \"...\"}"
}

---

4. Loading

import json

def read_jsonl(path):
    with open(path, encoding="utf-8") as f:
        for line in f:
            if line.strip():
                yield json.loads(line)

def to_qa(record):
    human = next(c["value"] for c in record["conversations"] if c["from"] == "human")
    gpt   = next(c["value"] for c in record["conversations"] if c["from"] == "gpt")
    return {"id": record["id"], "question": human, "answer": gpt, "source": record["source"]}

drugs = [to_qa(r) for r in read_jsonl("drugs_overdose_nepali_qa_.jsonl")]
hiv   = [to_qa(r) for r in read_jsonl("hiv_mental_health_nepali_qa_.jsonl")]
print(len(drugs), len(hiv))   # 800 800

4.1 Splitting without leakage

The same question appears in several records with different answers (Section 5.3). A random row split puts the same question in train and test. Split by question text instead:

import random

questions = sorted({r["question"] for r in drugs})
random.Random(42).shuffle(questions)
test_q = set(questions[: int(0.15 * len(questions))])
train = [r for r in drugs if r["question"] not in test_q]
test  = [r for r in drugs if r["question"] in test_q]

---

5. Diversity

Figures are from diversity_report.json (shared scripts/diversity_report.py) and from checks made while writing this README.

5.1 Overview

Metric	Drugs	HIV
Records	800	800
Unique questions	400 (50.0%)	380 (47.5%)
Unique answers	96 (12.0%)	156 (19.5%)
Distinct question pairs	800	800
Answers used for more than one question	95	150
Most-used answer (count)	14	13
Questions with more than one distinct answer	241	246
Max distinct answers for one question	6	7
Questions with an instruction prefix	91.0%	91.0%
Median question length (chars)	55	66
Median answer length (chars)	173	119
Questions containing Latin letters	7.5%	1.2%
Answers containing Latin letters	45.2%	17.6%

Every (question, answer) pair is unique, but each question and each answer is shared. This pattern looks like text split into chunks and paired with questions in several ways. Treat the file as a set of paired statements, not as a set of independent questions.

5.2 Answer types

Answer type	Drugs	HIV
Descriptive sentence	95.0%	72.4%
Number inside a sentence	0%	27.6% (all are the hotline number 1166)

The drugs file has no numbers at all. The HIV file has one number, the helpline.

5.3 Question-answer structure

The same question is paired with several different answers, and the same answer is paired with several questions. Two examples from the drugs file:

• "ओभरडोज भएको मान्छेलाई कसरी बचाउने?" (how to help a person who has overdosed) is paired with first-aid answers, which fit the question.
• "लागू औषध ओभरडोजको जोखिम कसरी कम गर्ने?" (how to reduce the risk of drug overdose) is paired with a definition of overdose, which fits less well.

In the HIV file, "HIV र मानसिक स्वास्थ्य बीचको सम्बन्ध के हो?" (what is the relationship between HIV and mental health) is paired with a helpline answer, which does not answer the question.

A lexical check (shared content words between question and answer) finds zero overlap in 15.9% of drugs records and 17.9% of HIV records. This is a weak signal. Nepali inflection hides real matches, so some zero-overlap pairs are valid paraphrases, and the check does not measure accuracy. Treat the percentages as a prompt for review, not as a count of errors.

5.4 Prefix and source markers

Most questions (91%) begin with a prompt prefix, and the prefixes repeat across records. In the drugs file, "पम्फलेट अनुसार" (according to the pamphlet) appears 90 times as a prefix. In the HIV file, "आधिकारिक जानकारी अनुसार" (according to official information) appears 86 times.

Answers also carry source markers: 219 drugs answers and 194 HIV answers end with "स्रोत: ..." (source), and 206 and 189 respectively contain a pamphlet or poster tag in brackets. These markers repeat the text and can be removed.

---

6. Content accuracy and framing (HIV file)

6.1 Yes-bias

This is the most serious problem in the collection.

Measure	Count
Answers starting with "हो" (yes)	317
Answers starting with "होइन" (no)	0
"हो" answers to questions that are not yes/no (for example, "what", "where", "which")	167

The HIV file never answers no: the count of answers starting with "होइन" is zero. Yes/no-style questions are answered with "हो" or with a statement, and the dataset never teaches a model to say no. Worse, 167 "yes" answers respond to "what" or "where" questions, which makes them meaningless. One example: "पोस्टर अनुसार, ११६६ नम्बर के को लागि हो?" (according to the poster, what is number 1166 for?) gets "हो, आत्महत्या रोकथाम हेल्प लाइन निःशुल्क छ" (yes, it is a free suicide-prevention helpline). The "yes" is not an answer to the question.

A model trained on this file will tend to agree with any framing. This matters most for mental health, where the wrong answer to "is this a sign of a problem?" can either alarm or reassure someone inappropriately.

6.2 Over-broad claims about symptoms

The HIV file treats many everyday experiences as signs of a mental health problem, in one-line answers with no qualification. Examples:

• "हो, चिडचिडापन, आक्रोश वा निराशा बढ्नु मानसिक स्वास्थ्य समस्याको संकेत हो।" (yes, increased irritability, anger or hopelessness is a sign of a mental health problem).
• "हो, भोक नलाग्ने वा निद्रा नलाग्ने समस्या मानसिक स्वास्थ्य समस्याको संकेत हुन सक्छ।" (yes, loss of appetite or sleep may be a sign of a mental health problem).
• "साथी वा परिवारसँग टाढा बस्ने बानी मानसिक स्वास्थ्य समस्याको संकेत हुन सक्छ।" (distancing from friends or family may be a sign of a mental health problem).

Each of these signs is real in some people, but none is diagnostic on its own, and the answers do not say so. A model trained on them may label ordinary tiredness or withdrawal as illness. Add the qualifiers that the source poster gives, or mark these as "may indicate" with a pointer to a clinician.

6.3 Unexplained acronyms

Key-population terms (PWID, MSM, TG, MLM) appear without expansion. A reader who does not know the terms cannot tell who the answer refers to. Expand them on first use or add a glossary.

6.4 Crisis information

The 1166 helpline appears in 221 answers. Two points to check:

• Verify the number against the official source before use. The file does not give its hours, its operator, or whether it is free to call from all networks.
• Repetition is not a substitute for safety design. A model that learns to give the number in every mental health answer may give it in answers where it is not relevant, and may not give it when a person is in danger and needs more than a number.

---

7. Limitations

1. Health content, not reviewed by a clinician. Do not use these answers as medical advice. Have a clinician or the publishing body review them before any deployment.
2. Paired statements, not independent questions. Section 5.1. Splitting by row leaks questions across train and test (Section 4.1).
3. Yes-bias in the HIV file. Section 6.1. No "no" answers exist.
4. Question-answer mismatch. Section 5.3. Some answers do not answer the question they are paired with.
5. Over-broad symptom claims. Section 6.2.
6. No "not in document" answers. Neither file teaches the model to say the material does not cover a question. Both files have 0 such answers.
7. Undated and unversioned. The pamphlet and poster editions are not recorded.
8. Attribution and license need checking. Section 1.1.
9. Repetitive answers. A few answers are used 13 to 14 times each.
10. No metadata variety. Every metadata value is the same within each file.

---

8. Recommended steps

Priority	Step	Reason
Critical	Have a clinician or NCASC review all answers before any use outside research.	Section 7, items 1 and 5.
Critical	Add "no" answers and "not covered in this material" answers to the HIV file, and remove the yes-bias.	Section 6.1.
High	Remove the "हो" from answers to wh-questions, and rewrite the 167 mismatched answers.	Section 6.1.
High	Verify the 1166 helpline (hours, operator, cost) against an official source, and add a date.	Section 6.4.
High	Add qualifiers to symptom answers ("may indicate", "not on its own").	Section 6.2.
High	Confirm the publisher and the license, and fix the attribution conflict.	Section 1.1.
High	Split by question, not by row.	Section 4.1.
Medium	Rewrite questions that are paired with unrelated answers.	Section 5.3.
Medium	Expand acronyms or add a glossary.	Section 6.3.
Medium	Remove the bracketed source tags and prefix markers from the answers if you want clean text.	Section 5.4.
Low	Record the edition and date of the source material.	Section 7, item 7.

---

9. Files

drugs_hiv_nepal/
├── README.md              This file
└── diversity_report.json  Shared diversity metrics for both files

The two data files are not copied into this folder. To regenerate the metrics:

python scripts/diversity_report.py --data_dir /path/to/files --out drugs_hiv_nepal/diversity_report.json

The shared script reads both files, since it processes every *.jsonl file in the directory. Move the other datasets out of the folder first, or pass a folder that contains only these two files.

WHO Health QA Dataset (Nepali, 1,300 rows)

A Nepali (Devanagari) factual question-answering dataset on health topics, built from seven public health documents published by the World Health Organization, the NCBI Bookshelf, and Harvard Health. Each record is a single-turn question with a short factual answer, stored as ShareGPT-style JSON Lines with per-row metadata.

Read Section 2 before using this dataset. About 46% of the rows (596 of 1,300) are numbered "point" questions whose answer does not depend on the number in the question. Exclude those rows from training and evaluation unless they are fixed.

---

Table of Contents

1. Overview
2. Critical: Natural vs Templated Rows
3. File and Schema
4. Source Breakdown
5. Content and Question Patterns
6. Diversity Analysis
7. Known Issues and Limitations
8. How to Load the Data
9. Recommended Usage
10. Licensing and Sources
11. Reproducing the Statistics
12. Changelog

---

1. Overview

Property	Value
File	WHO_Text_health_qa_1300_nepali.jsonl
Rows	1,300
Size	2.18 MB
MD5	c6861a3de0d21b6ee5306370d2106e40
Language	Nepali (ne, code npi)
Script	Devanagari, with some embedded Latin (abbreviations and technical terms)
Domain	Health (स्वास्थ्य)
Topics	Cholera, cervical screening (abnormal Pap smear), abortion care, alcohol withdrawal, 17-hydroxylase deficiency (CAH), elder abuse, healthy ageing
Question type (metadata)	Fact-based (तथ्यमा आधारित प्रश्न)
Behaviour label (metadata)	Provide informational answer (जानकारीमूलक उत्तर दिने)
Generation type	Real (from source documents)
License field	Apache-2.0 (permissive). See Section 10.
Unique questions	1,299 (one question appears twice, see Section 7)
Unique answers	339

The metadata matches the text: content_script says Devanagari, and the text is Devanagari. This differs from the romanized datasets in earlier releases, where the metadata did not match the text.

---

2. Critical: Natural vs Templated Rows

The file contains two kinds of rows. They differ in how the question was produced and in how useful they are.

Group	Row index	Rows	Unique answers	What the question looks like	Usable?
Natural	1 to 704 (health_qa_0001 to health_qa_0704)	704	332	A normal question about a specific fact (for example कलेरा कसरी फैलिन्छ?)	Yes
Templated	705 to 1,300 (health_qa_0705 to health_qa_1300)	596	7	A topic name plus a point number (for example १७-OH डेफिसिएन्सी व्यवस्थापन बिन्दु 705 के हो?)	No, see below

2.1 What the templated rows look like

Each templated row has a question of the form <topic> बिन्दु N के हो? (or तथ्य नम्बर N के हो?, or तथ्य N के हो?). The number N increases by one from row to row, and the seven sources take turns in a fixed cycle. Every source has exactly one fixed answer for all of its templated rows.

Source	Templated rows	The one answer used for every templated row of that source
ncbi_17_hydroxylase	86	ग्लुकोकोर्टिकोइड र सेक्स हर्मोन रिप्लेसमेन्ट जीवनभर आवश्यक।
harvard_alcohol_withdrawal	85	गम्भीर लक्षणमा अस्पताल/ICU उपचार र इमानदार इतिहास आवश्यक।
who_abortion_care	85	गर्भपतन सेवामा पहुँच, सुरक्षा र सम्मान स्वास्थ्य अधिकारको आधार हो। तथ्य सन्दर्भ: WHO abortion care।
who_cholera	85	सुरक्षित पानी, WASH, OCV र छिटो उपचार कलेरा नियन्त्रणका स्तम्भ हुन्।
who_elder_abuse	85	केयरगिभर समर्थन, जागरूकता र बहु-टोली प्रतिक्रिया प्रभावकारी रणनीति हुन्।
who_ageing	85	उमेर-अनुकूल समाज र लगानीले लामो स्वस्थ जीवन सम्भव बनाउँछ।
ncbi_abnormal_pap	85	नियमित Pap test र HPV भ्याक्सिन सर्भिकल क्यान्सर रोकथामका मुख्य उपाय हुन्।

2.2 Why these rows are a problem

• The answer does not respond to the question. The same answer is given whether the question asks about point 705 or point 1,300.
• A point number asks for one specific item. At most one of the 85 or 86 questions per source can have a correct answer, and the rest are wrong.
• Training on these rows teaches the model that a numbered question gets a generic answer, which is the opposite of what a QA model should learn.
• The question wording itself (बिन्दु 705 के हो?) does not appear in any natural question.

Recommendation: exclude the templated rows. The loader in Section 8 does this with --natural-only or load_natural(). If you need the numbered content, regenerate those rows from the source documents, with the actual point text as the answer.

2.3 Impact of excluding them

Measure	All rows	Natural only
Rows	1,300	704
Unique answers	339	332
Type-token ratio (questions)	0.1523	0.1667
Type-token ratio (answers)	0.1050	0.2360
Median answer words	8	6

Removing the templated rows roughly halves the dataset, but it more than doubles answer diversity (answer TTR goes from 0.105 to 0.236).

---

3. File and Schema

The file is JSON Lines, one object per line. It uses the same top-level schema as the other datasets in this collection.

3.1 Top-level fields

Field	Description
id	Row identifier, health_qa_0001 to health_qa_1300. The row number matches the id.
conversations	Two turns. from: "human" is the question, from: "gpt" is the answer.
source	Source URL.
source_name	One of seven source names (Section 4).
source_repo	Always World Health Organization / NCBI / Harvard Health (a combined label).
source_config, source_split, source_revision	Always default, train, v1.
source_row_id	Same as id.
language, language_code	Always ne and npi.
script	Always Deva. Accurate for this file.
license, license_tier	Always Apache-2.0 and permissive. See Section 10.
task_type	Always instruction-following.
generation_type, condition	Always real.
url	Source page URL. Seven distinct values (Section 4).
metadata_json	JSON-encoded string. Call json.loads on it.

3.2 Fields inside metadata_json

Key	Value
generation_domain	स्वास्थ्य (Health), all rows
generation_category	Varies by source (Section 4)
generation_sub_domain	Varies by source (Section 4)
behavior	जानकारीमूलक उत्तर दिने (Provide informational answer), all rows
behavior_definition	Keep the meaning of the question and the facts of the answer, and convert into pure Nepali. All rows.
question_type	तथ्यमा आधारित प्रश्न (fact-based question), all rows
content_language	नेपाली, all rows
content_script	देवनागरी, all rows
english_content_allowed	False, all rows

3.3 Example record (shortened)

{
  "id": "health_qa_0001",
  "conversations": [
    {"from": "human", "value": "<natural question in Nepali>"},
    {"from": "gpt",   "value": "<short factual answer in Nepali>"}
  ],
  "source_name": "who_cholera",
  "language_code": "npi",
  "script": "Deva",
  "metadata_json": "{\"generation_category\": \"संक्रामक रोग\", \"question_type\": \"तथ्यमा आधारित प्रश्न\", ...}"
}

---

4. Source Breakdown

Source (source_name)	Category (generation_category)	Sub-domain	Total	Natural	Templated	URL
who_cholera	संक्रामक रोग (Infectious disease)	कलेरा (Cholera)	264	179	85	WHO cholera fact sheet (254 rows); WHO cholera upsurge page (10 rows)
ncbi_abnormal_pap	स्त्री रोग (Gynaecology)	सर्भिकल स्क्रिनिङ (Cervical screening)	217	132	85	NCBI Bookshelf NBK560850
who_abortion_care	प्रजनन स्वास्थ्य (Reproductive health)	गर्भपतन सेवा (Abortion care)	193	108	85	WHO abortion fact sheet
harvard_alcohol_withdrawal	लत र विड्रअल (Addiction and withdrawal)	अल्कोहल विड्रअल (Alcohol withdrawal)	186	101	85	Harvard Health A to Z page
ncbi_17_hydroxylase	एन्डोक्राइनोलोजी (Endocrinology)	CAH 17-OH	168	82	86	NCBI Bookshelf NBK560850
who_elder_abuse	उमेर स्वास्थ्य (Ageing health)	वृद्ध दुर्व्यवहार (Elder abuse)	140	55	85	WHO elder abuse topic page
who_ageing	उमेर स्वास्थ्य (Ageing health)	स्वस्थ उमेर (Healthy ageing)	132	47	85	WHO ageing topic page
Total			1,300	704	596

Note: ncbi_17_hydroxylase and ncbi_abnormal_pap share the same NCBI book URL (NBK560850). Use source_name, not url, to separate them.

4.1 Per-source lexical diversity

These values include templated rows, because they are computed on all rows of each source. Use the natural-only figures for training decisions.

Source	Unique questions	Unique answers	TTR (questions)	TTR (answers)	Median question words	Median answer words
who_cholera	264	75	0.168	0.165	6	8
ncbi_abnormal_pap	216	63	0.196	0.161	6	8
who_abortion_care	193	53	0.186	0.120	7	12
harvard_alcohol_withdrawal	186	46	0.177	0.144	6	8
ncbi_17_hydroxylase	168	48	0.187	0.207	7	7
who_elder_abuse	140	28	0.198	0.117	7	9
who_ageing	132	27	0.200	0.093	7	9

---

5. Content and Question Patterns

The patterns below describe the 704 natural rows, since those are the rows to use.

5.1 Question words (natural rows)

Counts are for whitespace-separated words, so they are approximate. A question can contain more than one.

Question word	Meaning	Count
के	what	362
को	who / whose	128
किन	why	127
कति	how much / how many	91
कसरी	how	68
कस्तो	what kind / what state	21
कहिले	when	10
कुन	which	6
कहाँ	where	5
कसले	by whom	3

The mix of question words shows that the questions ask different kinds of things: definitions (के हो?), causes (किन?), quantities (कति?), and processes (कसरी?). This is more varied than the single "fact-based" label in the metadata suggests.

5.2 Common question openers (natural rows)

Opener	Count
गर्भपतन सेवामा (about abortion care)	21
वृद्ध दुर्व्यवहारमा (about elder abuse)	12
उमेर बढ्दा (as age increases)	12
१७-OH मा (about 17-OH)	12
अनियोजित गर्भावस्था (unintended pregnancy)	9
असुरक्षित गर्भपतन (unsafe abortion)	9
४६,XY महिला (46,XY female)	9
अल्कोहल विड्रअलमा (about alcohol withdrawal)	9
WHO स्वास्थ्य (WHO health)	8
अस्पताल जाने (going to hospital)	8

5.3 Answer features (natural rows)

Feature	Count	Share
Answers containing a digit	459 of 704	65.2%
Answers containing Latin text (abbreviations, terms)	263 of 704	37.4%
Questions containing Latin text	311 of 704	44.2%

Examples of embedded Latin: Vibrio cholerae, CYP17A1, Pap smear, HPV, WASH, ICU, OCV, shakes.

5.4 Example natural pairs

Source	Question	Answer
who_cholera	हालका वर्षहरूमा कलेराको अवस्था कस्तो छ?	दशकौंको प्रगतिको बाबजुद कलेरा विश्वव्यापी रूपमा पुनरुत्थान भएको छ, लामो समयसम्म केस नदेखिएका देशहरूमा समेत।
who_cholera	कलेरा कसरी फैलिन्छ?	कलेरा खाना र पानीमा Vibrio cholerae ब्याक्टेरियाको दूषितताबाट फैलिन्छ, प्रायः मलबाट।
ncbi_abnormal_pap	Pap smear कसले विकास गरे?	George Papanicolaou ले सर्भिकल साइटोलोजी स्क्रिनिङ विकास गरे, जसबाट Pap smear नाम आयो।
ncbi_abnormal_pap	Pap test ले के गर्छ?	यो सर्भिक्सको ट्रान्जिसनल जोनबाट सेल संकलन गरी precancerous परिवर्तन जाँच्छ।
ncbi_17_hydroxylase	CAH को सबैभन्दा सामान्य कारण के हो?	२१-हाइड्रोक्सिलेज डेफिसिएन्सी, जसले ९०% भन्दा बढी केस ओगट्छ।
harvard_alcohol_withdrawal	अल्कोहल विड्रअलका सामान्य लक्षण के हुन्?	काँप्ने (shakes), निद्रा नलाग्ने, चिन्ता र अन्य शारीरिक तथा मानसिक लक्षण।
who_ageing	२०३० सम्म ६० वर्ष माथिका मानिस कति हुने अनुमान छ?	२०३० सम्म १.४ बिलियन हुनेछ।
who_elder_abuse	समुदायमा कति प्रतिशत वृद्धले दुर्व्यवहार अनुभव गर्छन्?	लगभग ६ मध्ये १ जना ६० वर्ष माथिकाले विगत वर्षमा समुदायमा कुनै न कुनै दुर्व्यवहार अनुभव गरे।

---

6. Diversity Analysis

6.1 Method

All numbers in this section use one tokenizer: lowercase the text, remove punctuation (? . , । ! : ; ( ) ' " / - –), and split on whitespace. Do not use a regex \w+ tokenizer for Devanagari. It splits words at vowel signs and gives wrong results.

Type-token ratio (TTR) is distinct tokens divided by total tokens. Bigram ratio is distinct two-word sequences divided by total two-word sequences. Higher means more variety. These ratios depend on size, so compare groups of similar size with care.

6.2 Lexical diversity by group

Group	Rows	Unique Q	Unique A	Question vocab	Answer vocab	TTR (Q)	TTR (A)	Bigram ratio (Q)	Median Q words	Median A words
All rows	1,300	1,299	339	1,285	1,179	0.1523	0.1050	0.3645	7	8
Natural only	704	703	332	682	1,165	0.1667	0.2360	0.4109	6	6
Templated only	596	596	7	622	64	0.1432	0.0102	0.3245	7	9

The templated rows have almost no answer variety (7 unique answers, answer TTR 0.0102). They pull full-dataset answer diversity down sharply.

6.3 Behaviour diversity

Level	Finding
Metadata behavior	One label for all 1,300 rows
Metadata question_type	One label for all 1,300 rows
Actual question intent (natural rows)	At least five types: definition, cause, quantity, process, and statistic or comparison
Answer length	Short answers. Natural rows have a median of 6 words. Across all rows the range is 2 to 21 words.

The metadata shows one behaviour and one question type, but the question wording shows more variety (Section 5.1). For behavioural analysis, classify questions from their text.

6.4 Question pattern diversity

Pattern	Natural rows	Templated rows
Standard Nepali question word (के, किन, कति, कसरी, को, कस्तो, कहिले, कुन, कहाँ, कसले)	Most rows	Not present
Numbered point (बिन्दु N, तथ्य नम्बर N, तथ्य N)	None	All 596 rows
Ends with ?	704	596

6.5 Answer diversity: repeated answers

Repeated answers in the natural rows are mostly the same fact asked more than once. Three answers appear eight times each:

Times	Answer
8	WHO स्वास्थ्य परिभाषा पूर्ण शारीरिक, मानसिक र सामाजिक कल्याण हो।
8	अनियोजित गर्भावस्था संख्या १२१ मिलियन वार्षिक हो।
8	असुरक्षित गर्भपतन प्रतिशत लगभग आधा हो।

In total, 54 natural answers repeat. Before treating them as separate facts, check that the questions ask different things. The templated answers (Section 2.1) are a separate and more serious case of answer reuse.

6.6 Overall rating (qualitative)

Aspect	Natural rows	Templated rows
Question variety	Medium to high	Very low (one numbered template)
Answer variety	Medium	Very low (7 answers)
Topic coverage	7 sources, 7 topics	Same 7 topics
Usefulness for training	Usable with care	Not usable as-is

---

7. Known Issues and Limitations

#	Issue	Impact	Suggested handling
1	Templated rows (596, 45.8%): numbered point questions with one fixed answer per source (Section 2).	Wrong answers for almost every row. Teaches the model that numbered questions get generic answers.	Exclude with load_natural(), or regenerate from the source text.
2	Citation tail in answers: तथ्य सन्दर्भ: WHO abortion care। ends 85 answers. All 85 are in the templated abortion rows.	Source metadata leaked into answer text.	Excluded if you use natural rows. If you keep templated rows, strip the tail with the code in Section 8.8.
3	Duplicate question with different answers: ASCUS के हो? appears twice, both from ncbi_abnormal_pap. One answer is a full definition, the other a short statement.	Conflicting targets for the same question.	Keep one, or keep both only if the model should accept either.
4	Repeated natural answers: 54 natural answers repeat. Three repeat 8 times each.	Possible over-weighting of a few facts.	Check the questions behind the repeats. Deduplicate if they ask the same thing.
5	Time-sensitive statistics: examples include the number of people over 60 in 2019 (about 1 billion) and the projection for 2030 (about 1.4 billion).	These may be out of date.	Check the current source before relying on any number.
6	Not medical advice. The content comes from health documents, and some answers describe treatment.	Risk if used as clinical guidance.	Use for research, education, or model evaluation, and verify against current clinical guidance.
7	Licence field and source terms. The file says Apache-2.0, but the sources are WHO, NCBI, and Harvard Health.	The Apache-2.0 label does not confirm the rights of the source text.	Check each source's terms before redistribution (Section 10).
8	Small natural set: only 704 usable rows.	Too small to train from scratch.	Use for evaluation, few-shot examples, or as part of a fine-tuning mix.
9	Two cholera URLs: 254 rows point to the cholera fact sheet, 10 to the cholera upsurge page.	Source attribution differs within one source name.	Use the url field if attribution matters.
10	Mixed Latin terms: abbreviations and technical terms appear unchanged (for example CYP17A1, Vibrio cholerae).	Normal for code-mixed text.	No action needed.
11	Zero-width joiner: one row (health_qa_0086) contains U+200D.	It is a valid Devanagari joiner.	Keep it.
12	Answer correctness not fully verified.	Some errors may exist.	Spot-check a random sample against the source before production use.

---

8. How to Load the Data

All examples assume the file is at DATA_FILE. They were tested against the uploaded file.

8.1 Plain Python

import json

DATA_FILE = "/path/to/WHO_Text_health_qa_1300_nepali.jsonl"

with open(DATA_FILE, encoding="utf-8") as fh:
    rows = [json.loads(line) for line in fh if line.strip()]

print(len(rows))                                   # 1300

row = rows[0]
print(row["conversations"][0]["value"])            # question
print(row["conversations"][1]["value"])            # answer
meta = json.loads(row["metadata_json"])            # metadata is a JSON string
print(meta["generation_category"])

8.2 Using the included loader

load_who_health_qa.py (included with this README) flags templated rows and can drop them.

from load_who_health_qa import load_all, load_natural, summarize

all_rows = load_all("/path/to/WHO_Text_health_qa_1300_nepali.jsonl")
summarize(all_rows)

natural = load_natural("/path/to/WHO_Text_health_qa_1300_nepali.jsonl")
print(len(natural))                                # 704
print(natural[0]["instruction"], "->", natural[0]["output"])

Command line:

python load_who_health_qa.py --file /path/to/WHO_Text_health_qa_1300_nepali.jsonl
python load_who_health_qa.py --file /path/to/WHO_Text_health_qa_1300_nepali.jsonl --natural-only

Expected output with --natural-only:

Total records: 704
Templated (numbered point) rows: 0
Per source:
    179  who_cholera
    132  ncbi_abnormal_pap
    108  who_abortion_care
    101  harvard_alcohol_withdrawal
     82  ncbi_17_hydroxylase
     55  who_elder_abuse
     47  who_ageing

8.3 HuggingFace datasets

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files={"who": "/path/to/WHO_Text_health_qa_1300_nepali.jsonl"},
    split="who",
)
print(ds)           # 1300 rows

8.4 Flag templated rows and keep the natural ones

import re
from datasets import load_dataset

TEMPLATED_RE = re.compile(r"(?:बिन्दु|तथ्य नम्बर|तथ्य|जानकारी)\s*[०-९0-9]+\s*के हो\?")

ds = load_dataset("json", data_files={"who": "/path/to/WHO_Text_health_qa_1300_nepali.jsonl"}, split="who")
ds = ds.add_column(
    "templated",
    [bool(TEMPLATED_RE.search(r["conversations"][0]["value"])) for r in ds],
)

natural = ds.filter(lambda r: not r["templated"])
print(natural.num_rows)     # 704

8.5 Convert to instruction or chat format

def to_alpaca(row):
    return {
        "instruction": row["conversations"][0]["value"],
        "input": "",
        "output": row["conversations"][1]["value"],
    }

def to_chat(row):
    return {
        "messages": [
            {"role": "user", "content": row["conversations"][0]["value"]},
            {"role": "assistant", "content": row["conversations"][1]["value"]},
        ]
    }

alpaca = natural.map(to_alpaca, remove_columns=natural.column_names)
chat = natural.map(to_chat, remove_columns=natural.column_names)
print(alpaca[0])

8.6 Source-level split (leak-safe)

Questions from the same source document share facts, so splitting by source is safer than a random split.

sources = sorted(set(natural["source_name"]))
test_sources = {sources[1], sources[4]}     # ncbi_17_hydroxylase, who_ageing

train = natural.filter(lambda r: r["source_name"] not in test_sources)
test = natural.filter(lambda r: r["source_name"] in test_sources)
print(train.num_rows, test.num_rows)        # 575 129

8.7 Remove exact duplicate pairs

seen = set()
keep = []
for i, row in enumerate(natural):
    key = (row["conversations"][0]["value"], row["conversations"][1]["value"])
    if key not in seen:
        seen.add(key)
        keep.append(i)

natural_unique = natural.select(keep)
print(natural_unique.num_rows)     # 704 (the natural rows have no exact duplicates)

8.8 Strip the citation tail (only needed if you keep templated rows)

import re

TAIL = re.compile(r"\s*तथ्य सन्दर्भ:.*$")

def strip_tail(row):
    row["conversations"][1]["value"] = TAIL.sub("", row["conversations"][1]["value"])
    return row

cleaned = ds.map(strip_tail)

8.9 Pandas

import pandas as pd

df = pd.read_json("/path/to/WHO_Text_health_qa_1300_nepali.jsonl", lines=True)
df["question"] = df["conversations"].apply(lambda c: c[0]["value"])
df["answer"] = df["conversations"].apply(lambda c: c[1]["value"])
print(df[["question", "answer"]].head())

---

9. Recommended Usage

Goal	Recommendation
Train a Nepali health QA model	Use the 704 natural rows only. The set is small, so add other data.
Evaluate a Nepali health QA model	Use the source-level split (Section 8.6) on the natural rows. Report results per source.
Study how a model handles numbered questions	Use the templated rows only as a labelled negative test, for example to check whether a model gives a generic answer to a numbered question. Do not use them as training targets.
Study cross-source generalisation	Hold out one source at a time.
Study code-mixed Nepali	Use the natural rows, which contain Latin abbreviations and terms.

General advice:

• Do not use this dataset for clinical decisions.
• Report results per source, since topics and difficulty differ.
• Check answers against the source before publishing any evaluation numbers.

---

10. Licensing and Sources

Source	Publisher	Where to check terms
who_cholera, who_abortion_care, who_elder_abuse, who_ageing	World Health Organization	WHO publication and fact-sheet reuse terms on who.int
ncbi_abnormal_pap, ncbi_17_hydroxylase	NCBI Bookshelf (National Library of Medicine)	NCBI Bookshelf copyright and reuse notice
harvard_alcohol_withdrawal	Harvard Health Publishing	Harvard Health site terms

The license field says Apache-2.0 for this file. That label describes the QA pairs as provided. It does not change the rights of the source texts, which have their own terms. Check those terms before redistributing the dataset or publishing a model trained on it.

---

11. Reproducing the Statistics

This script uses the same tokenizer as Section 6.1.

import json, re, statistics

PUNCT = re.compile(r"[?.,।!:;()'\"/\-–]+")
TEMPLATED_RE = re.compile(r"(?:बिन्दु|तथ्य नम्बर|तथ्य|जानकारी)\s*[०-९0-9]+\s*के हो\?")

def tokens(text):
    return [t for t in PUNCT.sub(" ", text.lower()).split() if t]

rows = [json.loads(l) for l in open("WHO_Text_health_qa_1300_nepali.jsonl", encoding="utf-8") if l.strip()]

groups = {
    "all": rows,
    "natural": [r for r in rows if not TEMPLATED_RE.search(r["conversations"][0]["value"])],
    "templated": [r for r in rows if TEMPLATED_RE.search(r["conversations"][0]["value"])],
}

for name, data in groups.items():
    Q = [r["conversations"][0]["value"] for r in data]
    A = [r["conversations"][1]["value"] for r in data]
    qt = [t for q in Q for t in tokens(q)]
    at = [t for a in A for t in tokens(a)]
    bg = [tuple(tokens(q)[i:i+2]) for q in Q for i in range(len(tokens(q)) - 1)]
    print(f"{name:10s} rows={len(data)} uQ={len(set(Q))} uA={len(set(A))} "
          f"TTR_q={len(set(qt))/len(qt):.4f} TTR_a={len(set(at))/len(at):.4f} "
          f"bigram_q={len(set(bg))/len(bg):.4f} "
          f"medQ={statistics.median(len(q.split()) for q in Q)} "
          f"medA={statistics.median(len(a.split()) for a in A)}")

Expected output matches Section 6.2 (values rounded to four decimals).

---

12. Changelog

• v1 (current): Initial README for WHO_Text_health_qa_1300_nepali.jsonl. Identified the 596 templated rows with one fixed answer per source, the citation tail in 85 answers, and the duplicate ASCUS question. Added load_who_health_qa.py with a natural-only option.