Romanized Nepali Health Instruction Dataset (REAL_SCRAP_HEALTH_COMPILE_ROMANIZED_DATASET)

A Latin-script (romanized) version of the Nepali health instruction dataset REAL_SCRAP_HEALTH_COMPILE_DATASET. It contains the same 87,182 question–answer records, in the same row order, with the same IDs, sources and metadata. Only the question and answer text has been converted from Devanagari into Latin letters, so Nepali is written the way it is typed on a Latin keyboard, for example aarthik barsha 2072/73 ko lakshit janasankhya.

The file is a parallel copy, not a new set of records. Read Section 2 before combining the two files.

Read this first: Every question and answer is in Roman script except 9 questions and 2 answers, which still contain Devanagari. The metadata still labels the text as Devanagari (script = Deva), which is wrong for this file. Only about 3.7% of records come from official Nepali tables. About 96% are synthetic or model-generated, and about 65% are US county mortality data translated into romanized Nepali. See Section 3.

---

---

1. At a Glance

Property	Value
File	REAL_SCRAP_HEALTH_COMPILE_ROMANIZED_DATASET.jsonl
File size	~117 MB (116,608,129 bytes)
Records (lines)	87,182 (no malformed lines)
Unique question–answer pairs	86,602 (580 exact duplicate copies)
Unique question strings	86,597
Turns per record	Exactly 2 (human → gpt)
Script of question text	Latin (romanized); 87,173 questions have no Devanagari
Script of answer text	Latin (romanized); 87,180 answers have no Devanagari
Questions containing ASCII digits	84,094 (96.4%)
Records with metadata	86,595 (99.3%)
Records with missing metadata	587 (0.7%)
Distinct source blocks (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
Median question / answer length	155 / 149 characters
Longest answer	1,876 characters

---

2. Relationship to the Devanagari Dataset

This file is a row-aligned romanization of REAL_SCRAP_HEALTH_COMPILE_DATASET.jsonl. These were checked row by row:

Check	Result
Same record order and same id for every row	87,182 of 87,182 rows
Same source_name and source_row_id	87,182 of 87,182 rows
Same metadata_json content	87,182 of 87,182 rows
Question text identical to the Devanagari file	0 of 87,182 (all rewritten)
Answer text identical to the Devanagari file	0 of 87,182 (all rewritten)
Same numeric values in question and answer	87,182 of 87,182 rows

The last check compares the digits in each romanized row with the digits in the original row, after converting Nepali digits to ASCII. No numbers were lost or changed in the conversion.

Practical rules

• Use one file per experiment. Both files contain the same underlying records, so if you use the Devanagari file for training and the romanized file for testing (or the reverse), the same facts appear in both splits. Your evaluation will be inflated.
• Use the Devanagari file for Nepali-script tasks. Use the romanized file for tasks where users type Nepali in Latin letters, or for models that work better with Latin tokenisation.
• Row i in one file corresponds to row i in the other, so you can join them by line number if you need both.

---

3. What Is Real and What Is Not

The condition and generation_type fields are identical to the Devanagari file.

condition	Records	Share	generation_type	What it means
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- or scenario-based questions built by the compiler. Covers the Economist / WHO 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.
null	587	0.7%	null	No condition or source metadata.

Implications

• The real label is the compiler's label for the source data. Romanization does not change it, and it has not been verified record by record.
• The model-generated and synthetic records can contain generation errors, which are also present in the romanized text (see Section 12).
• The largest block (CDC, 64.75% of records) is about United States county and state mortality statistics. It is not Nepal data, even though it is written in romanized Nepali.

---

4. File Format and Field Reference

The file is JSON Lines: one JSON object per line, UTF-8 encoded. Each line ends with \r\n, so strip whitespace before json.loads.

4.1 Top-level fields

Field	Type	Always present?	Description
id	string	Yes	Record identifier. Not unique: 461 IDs appear twice, each time for a different question.
conversations	list of 2 objects	Yes	[{"from": "human", "value": ...}, {"from": "gpt", "value": ...}].
source, source_name, source_repo, source_config, source_split, source_revision, source_row_id	string or null	Mostly	Provenance fields, identical to the Devanagari file.
language	string or null	Mostly	ne (Nepali). This is the language, not the script.
language_code	string or null	Mostly	npi (ISO 639-3 for Nepali). Correct as a language code.
script	string or null	Mostly	Deva on all 86,595 tagged records. This is incorrect for this file, which is Latin-script. See Section 6.4.
license, license_tier	string	Yes	apache-2.0 and permissive.
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 or null	Mostly	Empty string for all tagged records.
metadata_json	string (JSON-encoded) or null	Mostly	A JSON string nested inside the record. Keys depend on the source block.

4.2 Keys inside metadata_json

The keys and values are identical to the Devanagari file. Labels such as behavior, question_type, generation_category and generation_domain are still in Devanagari or English, as in the original.

Key	Records	Meaning
domain	57,648	drug_poisoning_mortality (CDC) or health_awareness.
model	57,648	gemini-3.5-flash-lite or himalaya-gemma-4-e2b-it.
generation_number	57,648	Sequence number within the generation run.
generation_domain	28,947	Domain label (5 values).
generation_category	28,947	Category label (21 values).
generation_sub_domain	28,947	Sub-category label (64 values).
behavior	28,947	Target response behaviour (5 values). See Section 7.
behavior_definition	28,947	Definition of the behaviour (Nepali or English).
question_type	28,947	Question-type label (22 values).
content_language	28,947	नेपाली or English. Both are wrong for romanized text. See Section 6.4.
content_script	28,947	देवनागरी (28,691) or Latin (256). देवनागरी is wrong for this file.
english_content_allowed	28,947	Whether English content was permitted for the block.
question_length, response_length	25,602	Generation-time length constraints (text, not measured values).
maximum_response_sentences	25,602	1 (25,346) or 3 (256).
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 MCQ questions (always 4).

4.3 Example record (romanized)

{
  "id": "sg_f2091e39ce979c94cbdf0f24acd823dc",
  "conversations": [
    {"from": "human", "value": "aarthik barsha 2072/73 ko lakshit janasankhya deta anusar, jajarakot jillaka lagi prajanana umeraka (15-49 barsha) wiwahit mahilaharuko rekarda kati raheko chha? ka. 147637 kha. 60393 ga. 197536 gha. 39382"},
    {"from": "gpt", "value": "gha. 39382"}
  ],
  "source_name": "target_population_fy2072_73",
  "language": "ne",
  "language_code": "npi",
  "script": "Deva",
  "license": "apache-2.0",
  "task_type": "instruction-following",
  "generation_type": "synthetic",
  "condition": "synthetic",
  "url": ""
}

English gloss: "According to the fiscal year 2072/73 target population data, how many married women of reproductive age (15–49 years) are recorded for Jajarkot district? ka. 147637 kha. 60393 ga. 197536 gha. 39382" → "gha. 39382".

---

5. Composition by Source

Source blocks, record counts and generator models are identical to the Devanagari file. The medians below are measured on the romanized text, so they are longer than the Devanagari medians (Latin letters take more characters than Devanagari for the same content).

Source (source_name)	Records	Share	Condition	Generator model	Domain	Q median	A median
Centers for Disease Control and Prevention	56,448	64.75%	model-generated	gemini-3.5-flash-lite	drug_poisoning_mortality	152	160
The Economist / WHO Vaccination Data	24,597	28.21%	synthetic	—	स्वास्थ्य (health)	193	23
nepali_health_awareness_sft	1,200	1.38%	model-generated	himalaya-gemma-4-e2b-it	health_awareness	62	1,151
communicable_outbreaks_2073_74	986	1.13%	real	—	स्वास्थ्य (health)	71	48
target_population_fy2072_73	749	0.86%	synthetic	—	स्वास्थ्य जनसांख्यिकी	200	10
(missing)	587	0.67%	—	—	—	77	74
fchv_district_2072_73	402	0.46%	real	—	स्वास्थ्य (health)	63	52
cb_imnci_fchv_budget	300	0.34%	real	—	स्वास्थ्य (health)	88	88
dengue_positive_cases_fy2071_74	256	0.29%	synthetic	—	Public Health	104	98
shsp_enrollment	183	0.21%	real	—	स्वास्थ्य (health)	92	43
women_bank_mobile	176	0.20%	real	—	स्वास्थ्य तथा सामाजिक	82.5	81
nhrc_research	107	0.12%	real	—	स्वास्थ्य (health)	95	68
pharmacy_licensing_exam_faq	100	0.11%	synthetic	—	सार्वजनिक स्वास्थ्य परीक्षा तथ्याङ्क	51	36.5
dohs_province_health	93	0.11%	real	—	स्वास्थ्य (health)	78	78
hospital_cost	92	0.11%	real	—	स्वास्थ्य (health)	64.5	69
fchv_home_delivery	89	0.10%	real	—	स्वास्थ्य (health)	97	45
dohs_health_stats	80	0.09%	real	—	स्वास्थ्य (health)	92	90
dog_bite_rabies	74	0.08%	real	—	स्वास्थ्य (health)	65	62
hospital_employees	71	0.08%	real	—	स्वास्थ्य (health)	63	58
npafp	68	0.08%	real	—	स्वास्थ्य (health)	89	45
malaria_trend	68	0.08%	real	—	स्वास्थ्य (health)	67	66.5
snake_bite	55	0.06%	real	—	स्वास्थ्य (health)	63	63
diarrhoea_treatment	54	0.06%	real	—	स्वास्थ्य (health)	110	112
immunization	49	0.06%	real	—	स्वास्थ्य (health)	69	70
tb_notification	39	0.04%	real	—	स्वास्थ्य (health)	94	90
measles_rubella	34	0.04%	real	—	स्वास्थ्य (health)	40	38
diarrhoea_incidence	34	0.04%	real	—	स्वास्थ्य (health)	111	110
fp_modern	33	0.04%	real	—	स्वास्थ्य (health)	92	94
health_services_trend	29	0.03%	real	—	स्वास्थ्य (health)	60	64
ayurvedic_council	28	0.03%	real	—	स्वास्थ्य (health)	107	87
adolescent_abortion	26	0.03%	real	—	स्वास्थ्य (health)	77	84
adolescent_anc	25	0.03%	real	—	स्वास्थ्य (health)	114	114
phc_orc	18	0.02%	real	—	स्वास्थ्य (health)	91	93.5
pmtct_trend	16	0.02%	real	—	स्वास्थ्य (health)	76	75
patient_types	10	0.01%	real	—	स्वास्थ्य (health)	60.5	59.5
nurses_reg	6	0.01%	real	—	स्वास्थ्य (health)	50.5	51.5
Total	87,182	100%

Observations

• The top two blocks make up 92.96% of records. They are built from a small number of templates.
• There are 29 small real blocks (10 to 986 records each). They cover the same Nepali health topics as the Devanagari file.
• The health-awareness block is the only source with long answers (median 1,151 characters).

---

6. Romanization: Scheme, Consistency and Coverage

6.1 What the romanization looks like

The romanization is informal and phonetic. It is not a standard scheme such as ISO 15919 or the Harvard-Kyoto system. The file does not document which scheme was used, so the spellings below are the only guide. Examples from the file:

Devanagari (original)	Romanized (this file)	English
आर्थिक वर्ष २०७२/७३	aarthik barsha 2072/73	fiscal year 2072/73
लक्षित जनसङ्ख्या	lakshit janasankhya	target population
कभरेज	kabharej	coverage
बिसिजी	bisiji	BCG
डेङ्गु	dengu	dengue
प्रजनन उमेर	prajanana umera	reproductive age
वि.सं.	wi.san.	Vikram Samvat (year abbreviation)
तथ्यांक	tathyanka	statistics / data

Digits remain ASCII (2072/73, 15-49), while the Devanagari original used Nepali digits (२०७२/७३, १५-४९). Punctuation in the original (। for the full stop) is replaced by ..

6.2 Spelling consistency

Some words are spelled one way throughout the file. Others have several spellings in use.

Word	Spellings found	Counts (whole file)	Consistent?
aarthik / arthik	aarthik / arthik	3,882 / 0	Yes (one spelling)
barsha / barsa	barsha / barsa	24,709 / 0	Yes
kabharej / kabheraj	kabharej / kabheraj	26,364 / 0	Yes
dengu / dengue	dengu / dengue	724 / 0	Yes (one spelling)
swasthya / swastha	swasthya / swastha	16,029 / 1,966	No: both are used
sunsari / sunasari	sunsari / sunasari	2 / 44	No: sunasari dominates
kanchanpur / kanchanapur	kanchanpur / kanchanapur	1 / 44	No
dhanusa / dhanusha	dhanusa / dhanusha	1 / 43	No
dadeldhura / dadeladhura	dadeldhura / dadeladhura	Both used (23 / 17 question mentions)	No

Place names are the most inconsistent. A model trained on this file will see the same district written in two ways, and the spelling will not always match between a question and its options. See Section 12.

6.3 Coverage of Latin script

Measure	Count	Share
Questions with no Devanagari characters	87,173	100% of questions except 9
Questions still containing Devanagari	9	0.01%
Answers still containing Devanagari	2	0.002%
Answers with at least one Latin letter	87,179	99.997%

• The 9 questions with leftover Devanagari all come from the target_population_fy2072_73 block. They include Devanagari place names or phrases inside an otherwise romanized question.
• The 2 answers with leftover Devanagari both come from nepali_health_awareness_sft.
• Some questions contain a stray Devanagari nukta (़) inside a romanized word (for example, aankada़abhanda). These are the same 9 questions.

6.4 Metadata script labels are not updated

The romanization changed only the text. The metadata was copied unchanged from the Devanagari file, so it now contradicts the content.

Field	Value in file	Correct value for this file	Records affected
script	Deva	Latn (Latin)	86,595
content_script (metadata)	देवनागरी (Devanagari)	Latin	28,691
content_script (metadata)	Latin	Latin (correct here)	256
content_language (metadata)	नेपाली (Nepali)	Romanized Nepali	28,691
content_language (metadata)	English	Romanized Nepali	256
language_code	npi	npi (correct: the language is still Nepali)	86,595

Do not filter on script, content_script or content_language in this file. Filter on source_name, condition or behavior instead.

---

7. Behaviour Diversity

The behavior labels are identical to the Devanagari file. Each one is still written in Devanagari or English.

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 option from four that matches the data.
तथ्यमा आधारित उत्तर दिने	Give fact-grounded answers	3,245	3.7%	All real blocks	Keep the key fact intact and answer in plain 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 from the values given, in plain Nepali.
(no label)	—	58,235	66.8%	CDC, health awareness, missing-metadata rows	Not specified.

Answer behaviour in the romanized text

Answer format	Records	Share	Example
Free-text sentence	58,954	67.6%	aarthik barsha 2073/74 ma pradesh 1 ma bisiji kabharej 92% thiyo.
MCQ correct-answer sentence (sahi uttar: <letter>) <value>)	24,597	28.2%	sahi uttar: gha) 73.78%.
Short numeric answer	1,682	1.9%	110 × 481 = 52,910.
Long explanatory paragraph (300+ characters)	1,200	1.4%	General health-awareness explanation.
Letter plus option text (ka. <value>)	749	0.9%	ka. 96244

Note that the format counts are computed by rules that differ from the Devanagari file, so the numbers are close but not identical to that file's figures.

---

8. Question-Pattern Diversity

8.1 Question types (metadata question_type)

The 22 labels are identical to the Devanagari file. The table shows the 12 most common.

question_type	Records	Source	Description
cohort_size	3,509	Economist / WHO	Size of a target cohort under a scenario.
cancer_prevented	3,493	Economist / WHO	Number of cancer cases prevented.
deaths_prevented	3,485	Economist / WHO	Number of deaths prevented.
possible_cancer_cases	3,484	Economist / WHO	Possible cancer cases.
possible_cancer_deaths	3,473	Economist / WHO	Possible cancer deaths.
prevented_case_percentage	3,295	Economist / WHO	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	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 needing a short data-grounded answer.
प्रत्यक्ष संवादात्मक प्रश्न (direct conversational question)	100	Pharmacy exam FAQ	Conversational question in exam-prep style.

The remaining 10 labels (highest_*_country, lowest_*_country, total_*) cover 197 records, all in the Economist / WHO block. Examples: a question asking which country achieved the highest cancer prevention under a scenario, and a question asking for a total cohort across all countries.

8.2 Surface patterns in the question text

These counts use keyword matching on the romanized text. The keywords are different from those used for the Devanagari file, so the two files' figures are not directly comparable. They are approximate.

Pattern	Romanized keywords used	Records	Share
Contains a year or fiscal-year reference	wi.san., aarthik barsha, barsha, any four-digit number	83,432	95.7%
Asks for a count ("how many")	kati	71,160	81.6%
Scenario or counterfactual	paridrishya, yadi	24,978	28.7%
Multiple-choice (four options, ka. / ka) markers)	Option markers	25,346	29.1%
Comparison between entities	badhi (more), kam (less), barabar (equal), tulana (comparison)	1,558	1.8%
Explanatory or "how/why"	kina (why), kasari (how), ke ho (what is)	1,328	1.5%
Asks about change over time	prawritti (trend), pariwartan (change)	67	0.1%

The very low counts for change and comparison mean these question types are rare in the file. Many comparisons are phrased with badhi chha wa kam? ("is it more or less?"), which the keyword list does not fully capture.

8.3 Question length

Question length (characters)	Records	Share
Under 50	426	0.5%
50–99	4,273	4.9%
100–149	23,090	26.5%
150–199	49,703	57.0%
200 or more	9,690	11.1%

The romanized questions are longer than the Devanagari ones (median 155 characters against 136), because Latin letters take more characters than Devanagari for the same content. The question length is 150–199 characters for most records.

8.4 Template structure

The CDC and Economist / WHO blocks use a small set of sentence templates, with the entity, year and metric changed. Economist questions open with wi.san. 2021 to wi.san. 2030 (22,741 of the 24,597 questions in this block), and CDC questions open with a place name followed by a year.

Lexical variety is therefore low within those two blocks. Entities, years and numbers vary, but the sentence frames repeat.

8.5 Multiple-choice format

Format	Records	Example
Options inline, answer as sahi uttar: <letter>) <value>	24,597	ka) 0% kha) 80% ga) 70% gha) 48% → sahi uttar: ka) 0%.
Options inline, answer as <letter>. <value>	749	ka. 96244 kha. 31167 ga. 191972 gha. 171945 → ka. 96244

The options are numbers or district names of the same kind as the answer, which makes them useful as distractors. Check a sample before using them for evaluation.

---

9. Answer Diversity

9.1 Answer length (characters)

Answer length	Records	Share
Under 25	19,080	21.9%
25–74	8,926	10.2%
75–149	16,295	18.7%
150–499	41,682	47.8%
500 or more	1,199	1.4%

The longest answer is 1,876 characters, compared with 1,625 in the Devanagari file. Most answers are 150–499 characters because romanized Nepali needs more characters than Devanagari for the same words.

9.2 Answer style

• Restating style. Sentence answers repeat the place and metric from the question, then give the number, for example: jepharsana debhis parish jilla ko wiwaranaanusar 2012 ma janasankhya 31427 thiyo.
• Calculation style. Some pharmacy answers show the arithmetic, for example 110 × 481 = 52,910. (the Devanagari original is the same calculation).
• Bare-value style. About one in five answers is only a number or a short phrase.
• Explanatory style. The health-awareness answers are general-audience explanations on topics such as alcohol avoidance, seasonal health and daily-life wellbeing.

---

10. Topic and Domain Diversity

The topics, domains and categories are identical to the Devanagari file, so Section 9 of that README gives the full breakdown. The main points are repeated here.

10.1 Top-level domain (metadata generation_domain)

Domain (original label)	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

10.2 Categories (top 12 of 21)

Category (original label)	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

10.3 Health topics

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, measles and rubella	immunization, measles_rubella
Family planning	fp_modern
Maternal and adolescent health	adolescent_anc, adolescent_abortion, pmtct_trend
Rabies and 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 statistics	dohs_province_health, dohs_health_stats
General health awareness	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

---

11. Temporal and Geographic Coverage

11.1 Time references

Reference type	Values in the file	Notes
Fiscal year (Nepali official)	Nine distinct values, including 2073/74, 2072/73, 2071/72 (romanized digits)	Nepali fiscal years run from mid-July to mid-July.
Vikram Samvat abbreviation (wi.san.)	2021 to 2030	Used in the Economist / WHO block (24,442 occurrences).
Gregorian year	Mostly CDC years, such as 2003, 2004, 2005, 2007, 2008, 2014	The CDC block uses Gregorian years.

Calendar caveat (as in the Devanagari file). The Economist / WHO block writes its years as wi.san. (Vikram Samvat). The values 2021 to 2030 look like Gregorian projection years, so the label may be a calendar error. Confirm the calendar before using these years for date arithmetic.

11.2 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). It is an approximation, because some real questions may not name a location.

Nepali districts that appear in the romanized text include jajarakot, sunasari, kanchanapur, dhanusha, tanahu, pyuthan, dadeldhura and lamjung, with spellings varying as described in Section 6.2. For example, dadeldhura and dadeladhura both occur.

---

12. Data Quality and Known Issues

#	Issue	Records affected	Impact	Recommended handling
1	Metadata script and language labels not updated. script says Deva; content_script says देवनागरी for romanized text; content_language says नेपाली or English.	86,595 (script); 28,691 (content_script); 28,947 (content_language)	Filtering on these fields gives the wrong subset or the wrong script.	Do not filter on these fields. Use source_name or condition instead.
2	Leftover Devanagari text. Nine questions and two answers were not fully romanized.	9 questions, 2 answers	Mixed-script input, and the model may learn two scripts for the same task.	Romanize the remainder, or exclude these 11 records.
3	Inconsistent romanization of place names. The same district is spelled two ways.	sunsari/sunasari (2/44), kanchanpur/kanchanapur (1/44), dhanusa/dhanusha (1/43)	Exact-match evaluation and entity lookup will fail.	Normalise to one spelling per place, or use a fuzzy match.
4	Inconsistent romanization of common words.	swasthya/swastha (16,029/1,966)	Minor, but adds noise to vocabulary and tokenisation.	Normalise or accept as natural variation.
5	Answer text does not match the option text in the target-population MCQ block. The answer letter is correct, but the value differs.	11 spelling variants and 8 different texts (of 749)	Exact-match scoring will mark some correct answers wrong.	Match on the letter, not the text. Check the 8 different-text cases by hand (examples: manang vs manaph, lower vs kam).
6	Stray markup in one option (kam <>gha. badhi</>).	1	Looks broken if rendered.	Strip the markup.
7	Stray Devanagari nukta inside romanized words.	9 (same as issue 2)	Breaks word matching.	Remove the nukta.
8	Duplicate IDs. The same id is used for two different questions.	461 IDs (922 records)	Joins and lookups by id will be wrong.	Use a hash of the question and answer, or the row index.
9	Exact duplicate question–answer pairs.	580 extra copies	Inflates counts and can leak between splits.	Drop duplicates before splitting.
10	Answers that may not match the hypothetical in the question (observed in the pharmacy block, as in the Devanagari file).	Not measured across the block	Teaches terse or off-target answers.	Verify numbers against the question before training.
11	Template repetition in the CDC and Economist / WHO blocks.	About 81,000 records	Low structural diversity.	Subsample by source, or add paraphrased variants.
12	Calendar labelling of the Economist / WHO years (wi.san.).	24,442 occurrences	Date arithmetic will be wrong if the calendar is misread.	See Section 11.1.
13	Empty url field.	All tagged records	No source links are available from the file.	Cite using source and source_repo.

What was checked

• Structure: all 87,182 lines parse as JSON, and every record has two turns.
• Row alignment with the Devanagari file: IDs, sources, metadata and numeric values all match row by row.
• MCQ answers: all 24,597 Economist answers match their option text exactly. For the 749 target-population answers, 730 match exactly, 11 are spelling variants and 8 differ.
• Script: counts of Devanagari and Latin characters in questions and answers.
• Markup and nukta artifacts, and duplicate IDs and pairs.

What was not checked

• The factual correctness of the answers against the source tables. Spot checks found the issues above, so do not assume the numbers are correct without verification.
• Whether each romanized sentence reads naturally. The romanization was checked for consistency, not for fluency.

---

13. Loading the Dataset

The demo script load_romanized_dataset_demo.py in this folder contains all of the code below. It was tested against the file. It is the same loader as the Devanagari version, with the file name changed.

13.1 Requirements

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

13.2 Core loader (standard library, memory-friendly)

import json
from collections import Counter
from pathlib import Path

DATA_PATH = Path("REAL_SCRAP_HEALTH_COMPILE_ROMANIZED_DATASET.jsonl")


# 1. Stream records one at a time (recommended for a 117 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
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__":
    counts = Counter(rec.get("condition") for rec in stream_records())
    print("By condition:", dict(counts))

    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

13.3 pandas

import json
import pandas as pd

rows = []
with open("REAL_SCRAP_HEALTH_COMPILE_ROMANIZED_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))

13.4 Hugging Face datasets

from datasets import load_dataset

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

print(ds)                   # 87,182 rows, 19 columns
mcq = ds.filter(lambda r: r["source_name"] == "target_population_fy2072_73")
print(len(mcq))             # 749

metadata_json stays a string. Decode it with json.loads when you need the fields inside it.

13.5 Checking the romanized text

import re

DEVANAGARI = re.compile(r"[\u0900-\u097F]")

def is_fully_romanized(record):
    text = record["conversations"][0]["value"] + " " + record["conversations"][1]["value"]
    return not DEVANAGARI.search(text)

clean = [r for r in load_all() if is_fully_romanized(r)]
print(len(clean))           # 87,171 (drops the 11 records with leftover Devanagari)

---

14. Recommended Filters and Recipes

Goal	Filter or action
Only official-table data	condition == "real" (3,245 records)
Fully romanized text only	Drop the 11 records with leftover Devanagari (Section 13.5)
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)
Remove 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
Compare against the Devanagari file	Join by row number. Keep each row in one split only.
Remove missing-metadata rows	source_name is not None (drops 587)

Balancing. Two blocks make up 93% of records. For training, subsample the large blocks or weight the small real blocks more heavily.

---

15. Intended Use and Limitations

Intended use

• Research on Nepali language models that work in Latin script, including romanized chat and transliteration-based input.
• Studying how script choice affects tokenisation, and how well models transfer between Devanagari and romanized Nepali.
• Building romanized Nepali health evaluation sets, after verifying the answers you intend to use.

Limitations

• Not a clinical or medical reference. Numbers are not validated against source documents. Do not use the dataset for diagnosis, treatment or policy decisions without checking the sources.
• Informal romanization. There is no documented standard. The same word can appear in several spellings, and place names are inconsistent.
• Mostly synthetic. About 96% of records are model-generated or template-generated, and the romanized text inherits their errors.
• Not representative of Nepal. About two-thirds of the file is US data.
• Limited question diversity. Two blocks use a handful of templates, and most questions are 150–199 characters.
• Metadata does not describe the text. script, content_script and content_language describe the original Devanagari text, not this file.
• Calendar ambiguity. The wi.san. years may be Gregorian.
• Licensing. The apache-2.0 tag is applied at the dataset level. Check the upstream publisher terms before redistribution or commercial use.

---

16. How These Statistics Were Computed

Every number in this README was computed directly from the romanized JSONL file, except the row-alignment and numeric checks, which compared it with the Devanagari file row by row. The checks were:

• Parsing all 87,182 lines, which all parsed.
• Counting the values of each top-level and metadata_json field.
• Comparing IDs, sources, metadata and numeric tokens with the Devanagari file row by row.
• Counting Devanagari and Latin characters in questions and answers.
• Counting exact duplicate pairs and duplicate IDs.
• Measuring character lengths (Unicode code points).
• Validating MCQ answers against option text with a parser that handles multi-word options. The parser splits options on the ka), kha), ga) and gha) markers, then compares the answer value with the matching option. Mismatches were classified as spelling variants (similarity ratio of 0.8 or higher) or different text.
• Keyword pattern counts (Section 8.2). These are approximate.

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

---

Files in this folder

File	Purpose
README.md	This document.
load_romanized_dataset_demo.py	Tested loader script for the romanized file: streaming, flattening, filtering and chat-format conversion.

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

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

The romanized (Latin-script) version of the WHO health QA dataset. It contains the same 1,300 question-answer pairs as the Devanagari file WHO_Text_health_qa_1300_nepali.jsonl, with the same ids, in the same order. Each pair is a single-turn health question with a short factual answer, drawn from seven public documents by the World Health Organization, the NCBI Bookshelf, and Harvard Health.

Read Section 2 before using this dataset. Only 278 of the 1,300 rows (21%) are clean. The other 1,022 rows are either numbered rows, whose answers do not depend on the question, or shared-answer rows, whose answers often fit only one of several questions. Use the loader in Section 8 to select rows.

---

Table of Contents

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

---

1. Overview

Property	Value
File	WHO_Text_health_qa_1300_ROMANIZED.jsonl
Rows	1,300
Size	2.01 MB
MD5	26e5774fc04d04dee85e4487d38c78c2
Language	Nepali, romanized (Latin script). Language code npi.
Script in text	Latin only. No Devanagari characters.
Domain	Health
Topics	Cholera, cervical screening (abnormal Pap smear), abortion care, alcohol withdrawal, 17-hydroxylase deficiency (CAH), elder abuse, healthy ageing
Question type (metadata)	Fact-based
Behaviour label (metadata)	Provide informational answer
Generation type	Real (from source documents)
License field	Apache-2.0 (permissive). See Section 11.
Unique questions	1,299 (one question appears twice, see Section 9)
Unique answers	339
Parallel file	WHO_Text_health_qa_1300_nepali.jsonl (Devanagari, MD5 c6861a3de0d21b6ee5306370d2106e40)

---

2. Critical: Three Kinds of Rows

Every row falls into exactly one of three groups. The loader in Section 8 assigns the group automatically.

Group	Rows	Share	What the question looks like	What the answer looks like	Use?
numbered	596	45.8%	<topic> bindu N ke ho? (or number N, tathya N, janakari N)	One fixed answer for all rows of the same source	No
shared_answer	426	32.8%	A normal question, but one of several aspect questions about the same fact	The same answer as 7 other rows	Not as-is
clean	278	21.4%	A normal question about one fact	A unique answer	Yes

2.1 Numbered rows (596)

Rows 705 to 1,300 (health_qa_0705 to health_qa_1300). Each question names a point number, for example:

• 17-oh dephisiensi wyawasthapana bindu 705 ke ho?
• alkohal widraala suraksha bindu 706 ke ho?
• garbhapatana sewa sambandhi tathya number 707 ke ho?

Every source has one fixed answer for all of its numbered rows, so the answer does not depend on the number:

Source	Numbered rows	The one answer used for every numbered row (romanized)
ncbi_17_hydroxylase	86	glukokortikoida ra seksa harmona riplesamenta jiwanabhar aawashyaka.
harvard_alcohol_withdrawal	85	gambhir lakshanma aspatal /icu upachar ra imanadar itihas aawashyaka.
who_abortion_care	85	garbhapatana sewama pahuncha, suraksha ra sammana swasthya adhikarko aadhar ho. tathya sandarbha: who abortion care.
who_cholera	85	surakshit pani, wash, ocv ra chhito upachar kalera niyantranka stambha hun.
who_elder_abuse	85	keyaragibhar samarthana, jagarukata ra bahu-toli pratikriya prabhawakari rananiti hun.
who_ageing	85	umera-anukul samaj ra laganile lamo swastha jiwana sambhav banaunchha.
ncbi_abnormal_pap	85	niyamit pap test ra hpv bhyaksina sarbhikal cancer rokathamka mukhya upaya hun.

A numbered question asks for one specific item. At most one of the 85 or 86 questions per source can be answered correctly by the fixed answer. The rest are wrong, so these rows should not be used as training targets.

2.2 Shared-answer rows (426)

These are normal natural-language questions in the first 704 rows, but they come in groups. One fact is asked eight times with eight different aspect endings, and all eight share one answer:

Aspect ending (romanized)	Meaning
... kati ho?	how many / how much
... ke ho?	what is
... bare ke thaha chha?	what is known about
... ko mahattwa ke ho?	what is the importance of
... le ke sanket garcha?	what does it indicate
... kina mahattwapurna chha?	why is it important
... kasari sambandhit chha?	how is it related
... ko prabhav ke ho?	what is the effect of

Example group (all 8 rows have the answer lakshan awadhi 5 din ho.):

Row	Question	Answer	Does the answer fit the question?
health_qa_0669	lakshan awadhi kati ho?	lakshan awadhi 5 din ho.	Yes, a quantity
health_qa_0674	lakshan awadhi kina mahattwapurna chha?	lakshan awadhi 5 din ho.	No, the answer is a duration, not an importance

Another example: mutation sankhya le ke sanket garcha? (what does the mutation count indicate?) has the answer mutation sankhya 150+ ho. (mutation count is 150+). The answer gives a value, not an indication, so it does not fully match the question.

In total, 53 groups of eight rows (424 rows) and one pair of rows with the same answer (2 rows, see below) make up the 426 shared-answer rows. Because one answer is written for one fact, it can match only some of the eight aspect endings. The examples above show mismatches. The file has not been checked row by row, so treat every shared-answer row as unverified.

The 2-row group is 90% bhanda badhi. and is shared by two questions from different sources (hpv bhyaksinale kati pratishat hpv-related cancer rokna sakchha? and 46,xy madhye kati pratishat mahila rupma hurkainchhan?). Both questions ask for a percentage, but the answer still needs to be checked against the source.

2.3 Clean rows (278)

These are natural questions whose answer appears only once in the dataset. Examples (romanized, exact text):

Question	Answer
who le swasthyalai kasari paribhashit gareko chha?	who le swasthyalai purna sharirik, manasik ra samajik kalyanko awastha bhanera paribhashit gareko chha, jasma rog wa ashaktatako anupasthiti matra hoina.
gharapariwar aaya kshati kati ho?	gharapariwar aaya kshati 922 miliyana ameriki dalar ho.

2.4 Effect on the dataset

Measure	All rows	Clean rows only
Rows	1,300	278
Unique answers	339	278
Type-token ratio (answers)	0.1059	Not computed for the subset
Median answer words	8	Not computed for the subset

Using only the clean rows gives a small set, but every answer matches a unique fact and each question is used once.

---

3. File and Schema

The file is JSON Lines, one object per line. It uses the same top-level schema as the Devanagari file.

3.1 Top-level fields

Field	Description
id	health_qa_0001 to health_qa_1300. Same ids as the Devanagari file.
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	World Health Organization / NCBI / Harvard Health (a combined label).
source_config, source_split, source_revision	default, train, v1.
source_row_id	Same as id.
language, language_code	ne and npi.
script	Deva. This does not match the romanized text. See Section 9.
license, license_tier	Apache-2.0 and permissive. See Section 11.
task_type	instruction-following.
generation_type, condition	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 and facts of the answer, 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. This does not match the romanized text. See Section 9.
english_content_allowed	False, all rows

The metadata fields are written in Devanagari even though the text is romanized. Filter on source_name or id rather than on script or content_script.

3.3 Example record (shortened)

{
  "id": "health_qa_0001",
  "conversations": [
    {"from": "human", "value": "who le swasthyalai kasari paribhashit gareko chha?"},
    {"from": "gpt",   "value": "who le swasthyalai purna sharirik, manasik ra samajik kalyanko awastha bhanera paribhashit gareko chha, jasma rog wa ashaktatako anupasthiti matra hoina."}
  ],
  "source_name": "who_abortion_care",
  "script": "Deva",
  "metadata_json": "{\"generation_category\": \"प्रजनन स्वास्थ्य\", ...}"
}

---

4. Source Breakdown

Source (source_name)	Category (generation_category)	Total	Numbered	Shared-answer	Clean	URL
who_cholera	संक्रामक रोग (Infectious disease)	264	85	120	59	WHO cholera fact sheet (254 rows); WHO cholera upsurge page (10 rows)
ncbi_abnormal_pap	स्त्री रोग (Gynaecology)	217	85	81	51	NCBI Bookshelf NBK560850
who_abortion_care	प्रजनन स्वास्थ्य (Reproductive health)	193	85	64	44	WHO abortion fact sheet
harvard_alcohol_withdrawal	लत र विड्रअल (Addiction and withdrawal)	186	85	64	37	Harvard Health A to Z page
ncbi_17_hydroxylase	एन्डोक्राइनोलोजी (Endocrinology)	168	86	41	41	NCBI Bookshelf NBK560850
who_elder_abuse	उमेर स्वास्थ्य (Ageing health)	140	85	32	23	WHO elder abuse topic page
who_ageing	उमेर स्वास्थ्य (Ageing health)	132	85	24	23	WHO ageing topic page
Total		1,300	596	426	278

ncbi_17_hydroxylase and ncbi_abnormal_pap share the same NCBI book URL. Use source_name to tell them apart.

---

5. Romanization

5.1 Verified properties

Check	Result
Same row count as Devanagari file	1,300 and 1,300
Same id order	Yes
Same source_name and url for every row	Yes
Digits preserved	0 mismatches in questions and 0 in answers (Devanagari digits converted to Latin)
Devanagari characters left in text	0
Zero-width joiner (U+200D)	1 occurrence, in a word that is broken (see Section 9, item 4)

5.2 Spelling conventions

The romanization is a custom Latin scheme. It is not ISO 15919 or any standard transliteration. Within this file the spelling of common words is very consistent:

Word	Spelling used	Count	Alternative spelling	Count
aawashyaka (necessary)	aawashyaka	176	awashyaka	0
chha (is)	chha	231	cha	0
ho (is)	ho	1,373	hou	0
wriddha (elderly)	wriddha	112	vriddha	0
upachar (treatment)	upachar	198	upchar	0
kalera (cholera)	kalera	196	kolera	0
aspatal (hospital)	aspatal	104	aspataal	0
garbhapatana (abortion)	garbhapatana	230	garbhapatan	0

Proper nouns and English technical terms are kept as written in the source, for example pap test, hpv, wash, ocv, icu, colposcopy.

5.3 Side-by-side with the Devanagari file

Devanagari question	Romanized question
WHO ले स्वास्थ्यलाई कसरी परिभाषित गरेको छ?	who le swasthyalai kasari paribhashit gareko chha?
प्रत्येक वर्ष कति गर्भावस्था अनियोजित हुन्छन्?	pratyek barsha kati garbhawastha aniyojit hunchhan?
घरपरिवार आय क्षति कति हो?	gharapariwar aaya kshati kati ho?
mutation संख्या ले के संकेत गर्छ?	mutation sankhya le ke sanket garcha?

---

6. Content and Question Patterns

The patterns below describe the 704 natural rows (first 704 rows), since the numbered rows follow a fixed template.

6.1 Question words

Counts are for whitespace-separated words. They come from the Devanagari text and apply to the romanized text one-to-one.

Question word (Devanagari)	Romanized form	Meaning	Count
के	ke	what	362
को	ko	who / whose / of	128
किन	kina	why	127
कति	kati	how much / how many	91
कसरी	kasari	how	68
कस्तो	kasto	what kind	21
कहिले	kahile	when	10
कुन	kun	which	6
कहाँ	kaha	where	5
कसले	kasle	by whom	3

6.2 Example natural pairs

Question	Answer
who le swasthyalai kasari paribhashit gareko chha?	who le swasthyalai purna sharirik, manasik ra samajik kalyanko awastha bhanera paribhashit gareko chha, jasma rog wa ashaktatako anupasthiti matra hoina.
gharapariwar aaya kshati kati ho?	gharapariwar aaya kshati 922 miliyana ameriki dalar ho.
mutation sankhya le ke sanket garcha?	mutation sankhya 150+ ho. (This is a shared-answer row, see Section 2.2)

The answers keep Latin-script Nepali with English terms left as written, for example WASH, HPV, ICU, Pap test.

---

7. Diversity Analysis

7.1 Method

All numbers use one tokenizer: lowercase, remove punctuation (? . , ! : ; ( ) ' " / - –), split on whitespace. Do not use a regex \w+ tokenizer on Devanagari text. This README uses the same method as the Devanagari README, so the numbers can be compared directly.

Type-token ratio (TTR) is distinct tokens divided by total tokens. Higher means more variety.

7.2 Lexical diversity: romanized vs Devanagari

Group	Version	Rows	Unique Q	Unique A	Question vocab	Answer vocab	TTR (Q)	TTR (A)	Bigram ratio (Q)
All	Romanized	1,300	1,299	339	1,279	1,190	0.1516	0.1059	0.3641
All	Devanagari	1,300	1,299	339	1,285	1,179	0.1523	0.1050	0.3645
Natural	Romanized	704	703	332	676	1,176	0.1652	0.2382	0.4100
Natural	Devanagari	704	703	332	682	1,165	0.1667	0.2360	0.4109
Numbered	Romanized	596	596	7	622	64	0.1432	0.0102	0.3245
Numbered	Devanagari	596	596	7	622	64	0.1432	0.0102	0.3245

The two versions have almost the same diversity. Romanization does not reduce variety in any meaningful way. The large gap between natural and numbered rows is caused by the numbered template (Section 2.1).

7.3 Behaviour diversity

Level	Finding
Metadata behavior	One label for all rows
Metadata question_type	One label for all rows
Actual question intent (natural rows)	At least five types: definition, cause, quantity, process, and statistic or comparison
Answer length	Median 8 words across all rows (6 for natural rows), range 2 to 21

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

7.4 Answer diversity

The answer diversity is low in the full file because of the numbered and shared-answer rows. Among the natural rows, 54 distinct answers are shared by 426 rows (60.5%). The 53 eight-row groups account for most of this (Section 2.2). Only 278 natural rows have unique answers.

---

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

8.2 Using the included loader

load_who_health_qa_romanized.py (included with this README) assigns each row to numbered, shared_answer, or clean.

from load_who_health_qa_romanized import load_all, load_clean, summarize

F = "/path/to/WHO_Text_health_qa_1300_ROMANIZED.jsonl"

summarize(load_all(F))              # counts per row type and per source
clean = load_clean(F)
print(len(clean))                   # 278
print(clean[0]["instruction"], "->", clean[0]["output"])

Command line:

python load_who_health_qa_romanized.py --file /path/to/WHO_Text_health_qa_1300_ROMANIZED.jsonl
python load_who_health_qa_romanized.py --file /path/to/WHO_Text_health_qa_1300_ROMANIZED.jsonl --clean-only --strip-tail

Expected output of the first command:

Total records: 1300
    596  numbered
    426  shared_answer
    278  clean
Per source (clean rows):
     59  who_cholera
     51  ncbi_abnormal_pap
     44  who_abortion_care
     41  ncbi_17_hydroxylase
     37  harvard_alcohol_withdrawal
     23  who_elder_abuse
     23  who_ageing

8.3 HuggingFace datasets

import re
from collections import Counter
from datasets import load_dataset

NUMBERED_RE = re.compile(r"(?:bindu|number|tathya|janakari)\s+\d+\s+ke ho\?\s*$", re.IGNORECASE)

ds = load_dataset("json", data_files={"who": "/path/to/WHO_Text_health_qa_1300_ROMANIZED.jsonl"}, split="who")
ds = ds.add_column("numbered", [bool(NUMBERED_RE.search(r["conversations"][0]["value"])) for r in ds])
print(Counter(ds["numbered"]))      # False: 704, True: 596

8.4 Keep only clean rows

from collections import Counter

natural = ds.filter(lambda r: not r["numbered"])
answer_count = Counter(r["conversations"][1]["value"] for r in natural)
clean = natural.filter(lambda r: answer_count[r["conversations"][1]["value"]] == 1)
print(clean.num_rows)               # 278

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

8.6 Source-level split (leak-safe)

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

train = clean.filter(lambda r: r["source_name"] not in test_sources)
test = clean.filter(lambda r: r["source_name"] in test_sources)
print(train.num_rows, test.num_rows)        # 214 64

8.7 Strip the citation tail

The 85 numbered rows from who_abortion_care end with tathya sandarbha: who abortion care.. The clean rows do not contain it, but strip it if you keep other rows.

import re

TAIL = re.compile(r"\s*tathya sandarbha:.*$", re.IGNORECASE | re.DOTALL)

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

cleaned = ds.map(strip_tail)

8.8 Fix the zero-width joiner in one answer

def fix_zwj(row):
    row["conversations"][1]["value"] = row["conversations"][1]["value"].replace("\u200d", "")
    return row

ds = ds.map(fix_zwj)

8.9 Pandas

import pandas as pd

df = pd.read_json("/path/to/WHO_Text_health_qa_1300_ROMANIZED.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. Known Issues and Limitations

#	Issue	Impact	Suggested handling
1	Numbered rows (596, 45.8%): point-number questions with one fixed answer per source (Section 2.1).	Wrong answers for almost every row. Teaches the model that numbered questions get generic answers.	Exclude. Use load_clean().
2	Shared-answer rows (426, 32.8%): 53 groups where one answer is given to eight aspect questions about the same fact (Section 2.2).	Seven of eight questions in most groups are answered with a statement about a different aspect.	Exclude. Keep at most one question per group if you need some of them, after checking the match manually.
3	Duplicate question: ascus ke ho? appears twice. One answer is a full definition (health_qa_0131). The other is the generic ASCUS answer used in the shared-answer group (health_qa_0542).	Conflicting targets for the same question.	Keep the full definition only.
4	Zero-width joiner inside a word: pur\u200dyaune in health_qa_0086.	Splits one word into two tokens and breaks the match with the Devanagari text. It is a romanization artifact, not a Devanagari conjunct.	Remove it (Section 8.8).
5	Citation tail: tathya sandarbha: who abortion care. at the end of 85 answers, all in numbered rows.	Source metadata leaked into the answer text.	Excluded with the numbered rows. Strip with Section 8.7 otherwise.
6	Metadata does not match the text: script (Deva), content_script (देवनागरी), and content_language (नेपाली) describe Devanagari, but the text is romanized.	Filtering on these fields returns the wrong set.	Filter on source_name, id, or the text. Do not use script.
7	Time-sensitive statistics: examples include the number of people over 60 in 2019 and the 2030 projection.	May be out of date.	Check the current source before relying on any number.
8	Not medical advice. Some answers describe treatment.	Risk if used as clinical guidance.	Use for research, education, or evaluation only.
9	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 (Section 11).
10	Small clean set: 278 rows.	Too small to train from scratch.	Use for evaluation, few-shot examples, or as part of a larger mix.
11	Custom romanization: not a standard transliteration. Names and terms are spelled by sound.	Names may not match official or English spellings.	Do not use romanized names as lookup keys without normalization.
12	Answer correctness not fully verified.	Some errors may exist.	Spot-check a random sample against the source before production use.

---

10. Recommended Usage

Goal	Recommendation
Train a romanized Nepali health QA model	Use the 278 clean rows. Add other data, because the set is small.
Evaluate a romanized model	Use the source-level split (Section 8.6) on the clean rows. Report per source.
Compare scripts	Use the clean rows from this file and the matching ids from the Devanagari file. The question and answer content is the same, so the only difference is the script.
Test robustness to aspect phrasing	Do not use the shared-answer groups as they are. Build a new labelled set in which each aspect question has its own correct answer.
Test handling of numbered questions	Use the numbered rows only as a labelled negative test. Do not use them as training targets.

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 evaluation numbers.

---

11. 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. 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 before redistributing the dataset or publishing a model trained on it.

---

12. Reproducing the Statistics

This script uses the tokenizer from Section 7.1.

import json, re, statistics
from collections import Counter

PUNCT = re.compile(r"[?.,!:;()'\"/\-–]+")
NUMBERED_RE = re.compile(r"(?:bindu|number|tathya|janakari)\s+\d+\s+ke ho\?\s*$", re.IGNORECASE)

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_ROMANIZED.jsonl", encoding="utf-8") if l.strip()]
numbered = [r for r in rows if NUMBERED_RE.search(r["conversations"][0]["value"])]
natural = [r for r in rows if not NUMBERED_RE.search(r["conversations"][0]["value"])]
answer_count = Counter(r["conversations"][1]["value"] for r in natural)
clean = [r for r in natural if answer_count[r["conversations"][1]["value"]] == 1]

for name, data in [("all", rows), ("natural", natural), ("numbered", numbered), ("clean", clean)]:
    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:9s} 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: the rows and unique counts match Section 2.4. TTR values match Section 7.2 for the romanized rows.

---

13. Changelog

• v1 (current): Initial README for the romanized WHO file. Verified alignment with the Devanagari file (ids, sources, digits). Found that the romanized file has the same numbered-row problem as the Devanagari file, and identified a second problem not described in the earlier Devanagari README: 53 groups of eight aspect questions that share one answer (424 rows). The earlier README described these as 54 repeated answers without giving the row count (426 rows, 60.5% of natural rows), and it did not identify the aspect pattern. Found a zero-width joiner inside a romanized word and a duplicate ascus ke ho? question. Added load_who_health_qa_romanized.py with three row types.

Healthcare Patient-Record QA Dataset (Romanized Nepali, 50,102 rows)

A large romanized (Latin-script) Nepali question-answering dataset built from patient-record fields. Each row is one question about one field of one patient record, with a short factual answer. The records come from the Kaggle dataset prasad22/healthcare-dataset.

Read Section 2 before using this dataset. Patient names are not unique. 20,247 rows (40.4%) belong to names that appear in more than one record, so their questions do not identify which record they refer to. The loader in Section 8 flags these rows.

---

Table of Contents

1. Overview
2. Critical: Patient Names Are Not Unique
3. File and Schema
4. Column 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, Privacy, and Sources
11. Reproducing the Statistics
12. Changelog

---

1. Overview

Property	Value
File	healthcare_romanized_qa_final_unique.jsonl
Rows	50,102
Size	75.25 MB
MD5	368cac7e6da36ee0af0859c30c153267
Language	Nepali, romanized (Latin script). Language code npi.
Script in text	Latin only. No Devanagari characters.
Domain	Healthcare, patient records (hospital admission and billing)
Source	Kaggle prasad22/healthcare-dataset (https://www.kaggle.com/datasets/prasad22/healthcare-dataset)
Columns used	14 of the 15 source columns. Name is used as the subject of each question, not as a question column.
Question type (metadata)	Fact-based
Behaviour label (metadata)	Provide factual answer derived only from the dataset
Generation type	Real (derived from source records by templates)
License field	Apache-2.0 (permissive). See Section 10.
Unique questions	50,102 (no exact duplicate questions)
Unique answers	50,102 (no exact duplicate answers)
Distinct patient names	38,163

The metadata fields script, content_script, and behavior_definition describe Devanagari text, but the text is romanized. See Section 7.

---

2. Critical: Patient Names Are Not Unique

Each question starts with a patient name, and each answer repeats that name. Many names appear in more than one record.

Measure	Value
Rows	50,102
Distinct names	38,163
Names that appear in 2 or more rows	8,308
Rows whose name appears in 2 or more rows (ambiguous)	20,247 (40.4%)
Rows whose name appears only once (unambiguous)	29,855 (59.6%)
Most frequent name	michael williams (13 rows)
Other frequent names	robert smith (12), david smith (12), michael smith (11), matthew smith (11)

2.1 Why this matters

A question such as michael williams ko umera kati barsha ho? (how old is michael williams?) has a single answer in the file, but 13 different patient records share that name. The question does not say which record it means. This causes three problems:

• The answer may be correct for one record and wrong for the others. The file does not say which.
• A model trained on these rows learns name-to-value links that are not reliable.
• An evaluation on these rows cannot be scored fairly, because the question has more than one valid answer.

Name ambiguity affects every column. The most affected columns (ambiguous rows) are Gender (1,507), Billing Amount (1,465), and Age (1,459).

2.2 What to do

• Use the unambiguous rows only (29,855 rows, loader option load_unambiguous).
• Or keep the ambiguous rows only for tasks where the name is not the point, and state this in the model card.
• Do not evaluate on the ambiguous rows.

---

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	healthcare_faq_NNNNN_CC. NNNNN is the source record position (0 to 55,497). CC is NNNNN mod 15, which is the column slot (0 to 13).
conversations	Two turns. from: "human" is the question, from: "gpt" is the answer.
source	Kaggle dataset URL.
source_name	Always healthcare_dataset.
source_repo	Always prasad22/healthcare-dataset.
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. This does not match the romanized text. See Section 7.
license, license_tier	Always Apache-2.0 and permissive. See Section 10.
task_type	Always instruction-following.
generation_type, condition	Always real.
url	Kaggle page URL.
metadata_json	JSON-encoded string. Call json.loads on it.

3.2 Fields inside metadata_json

Key	Value
generation_domain	Healthcare, all rows
generation_category	Patient Medical Records, all rows
generation_sub_domain	Hospital Admission & Billing, all rows
column_focus	The source column asked about (14 values, Section 4)
original_row_index	Source record position, 0 to 55,497. Distinct for every row.
behavior	Provide factual answer derived only from the dataset, all rows
behavior_definition	Devanagari text: keep the meaning of the question and the facts of the answer, and convert into pure Nepali. See Section 7.
question_type	fact-based, all rows
content_language	नेपाली, all rows
content_script	देवनागरी, all rows. See Section 7.
english_content_allowed	False, all rows

3.3 Example record (shortened)

{
  "id": "healthcare_faq_00000_00",
  "conversations": [
    {"from": "human", "value": "bobby jackson ko umera kati barsha ho?"},
    {"from": "gpt",   "value": "bobby jackson ko umera 30 barsha ho."}
  ],
  "source_name": "healthcare_dataset",
  "language_code": "npi",
  "script": "Deva",
  "metadata_json": "{\"column_focus\": \"Age\", \"original_row_index\": \"0\", ...}"
}

---

4. Column Breakdown

Each source record contributes one row. The row's column_focus says which field was asked about.

column_focus	Rows	Answer type	Example answer
Age	3,593	Integer (years)	bobby jackson ko umera 30 barsha ho.
Test Results	3,590	Category	robert bauer ko parikshan natija samanya (normal) aayo.
Hospital	3,589	Name (text)	edward edwards group middleton aspatalma bharna bhaeka thie.
Discharge Date	3,588	Date (YYYY-MM-DD)	aaron martinez 2023-09-05 ma discharja bhaeka thie.
Gender	3,584	Category	leslie terry purush hun.
Medical Condition	3,584	Category with English term in brackets	andrew watts lai madhumeha (diabetes) ko samasya chha.
Medication	3,578	Category	connor hansen lai penisilina aushadhi diieko thiyo.
Blood Type	3,574	Category	danny smith ko rakta samuha a- ho.
Insurance Provider	3,574	Category	christina martinez ko bima pradayak cigna ho.
Billing Amount	3,572	Number (decimal)	jasmine aguilar ko biling rakam lagabhag 50119.22 thiyo.
Date of Admission	3,571	Date (YYYY-MM-DD)	adrienne bell 2022-09-19 ma aspatal bharna bhaeka thie.
Room Number	3,571	Integer	christopher berg lai kotha number 249 ma rakhieko thiyo.
Admission Type	3,568	Category	michelle daniels ko bharna prakar atyawashyak (urgent) thiyo.
Doctor	3,566	Name (text)	emily johnson ko upachar doctor taylor newton le gareka thie.
Total	50,102

Each source record gives at most one row. The column asked about is fixed by the record position: column = record_position mod 15. Slot 14 would be the Name field, which is used as the subject and never asked, so 3,700 positions (one in every 15) produce no row. Of the 51,800 remaining positions (out of 55,500), 50,102 are present, so 1,698 are absent. The file name says unique, but the data does not say why these were removed.

4.1 Categorical values (from the answers)

Column	Value	Count
Gender	mahila (female)	1,800
Gender	purush (male)	1,784
Blood Type	b-	478
Blood Type	a-	470
Blood Type	o+	463
Blood Type	b+	449
Blood Type	ab-	443
Blood Type	a+	428
Blood Type	ab+	427
Blood Type	o-	416
Medical Condition	cancer	632
Medical Condition	dam (asthma)	612
Medical Condition	gathiya (arthritis)	607
Medical Condition	madhumeha (diabetes)	582
Medical Condition	uchcha raktachap (hypertension)	578
Medical Condition	motopana (obesity)	573
Insurance Provider	cigna	730
Insurance Provider	medicare	723
Insurance Provider	aetna	722
Insurance Provider	unitedhealthcare	704
Insurance Provider	blue cross	695
Admission Type	atyawashyak (urgent)	1,197
Admission Type	aakasmik (emergency)	1,189
Admission Type	nirwachit (elective)	1,182
Test Results	anirnayak (inconclusive)	1,227
Test Results	asamanya (abnormal)	1,208
Test Results	samanya (normal)	1,155
Medication	espirina	750
Medication	aaibuprophena	730
Medication	lipitar	722
Medication	penisilina	709
Medication	pyarasitamol	667

The category values are balanced. Medication names are romanized from their pronunciation (for example espirina for aspirin, pyarasitamol for paracetamol).

4.2 Numeric ranges

Column	Min	Max	Median	Notes
Age	13	88	53	3,593 rows
Billing Amount	51.33	52,373.03	about 25,833	3,566 rows parsed; 6 rows did not match the expected pattern
Room Number	101	500	not computed	399 distinct room numbers
Date of Admission	2019-05-09	2024-05-07	not computed	3,571 rows
Discharge Date	2019-05-13	2024-06-01	not computed	3,588 rows

Every billing answer uses the word lagabhag (approximately), even though the value is a specific number (see Section 7).

---

5. Content and Question Patterns

5.1 Question templates

Each question has one of 14 fixed templates. The patient name is the only variable part (the name is placed before the template text). Replacing each name with <N> leaves exactly 14 distinct question templates:

Column	Question template (name replaced by <N>)
Age	<N> ko umera kati barsha ho?
Gender	<N> ko linga ke ho?
Blood Type	<N> ko rakta samuha ke ho?
Medical Condition	<N> lai kuna rog/chikitsa awastha chha?
Date of Admission	<N> kahile aspatal bharna bhaeka thie?
Doctor	<N> ko upachar garne doctor ko hun?
Hospital	<N> kuna aspatalma bharna bhaeka thie?
Insurance Provider	<N> ko bima pradayak kuna ho?
Billing Amount	<N> ko biling rakam kati thiyo?
Room Number	<N> kuna kotha numberma rakhieka thie?
Admission Type	<N> ko bharna prakar kasto thiyo?
Discharge Date	<N> kahile discharja bhaeka thie?
Medication	<N> lai kuna aushadhi diieko thiyo?
Test Results	<N> ko parikshan natija ke aayo?

5.2 Answer patterns

The answers follow the same column logic as the questions. After replacing the name, there are 14,115 distinct answer strings, because each answer also contains a different value (age, date, room number, amount, and so on). The answers are short: median 8 words (range 4 to 12). Every question ends with ? and every answer ends with ..

5.3 Name share of tokens

Names make up 29.0% of all question tokens. Most of the lexical content in the questions is the name, not the question wording.

---

6. Diversity Analysis

6.1 Method

All numbers use one tokenizer: lowercase, remove punctuation (? . , ! : ; ( ) ' " / - –), split on whitespace. Type-token ratio (TTR) is distinct tokens divided by total tokens. Bigram ratio is distinct two-word sequences divided by total two-word sequences.

6.2 Lexical diversity

Group	Rows	Question vocab	Answer vocab	TTR (Q)	TTR (A)	Bigram ratio (Q)	Median Q words	Median A words
All rows	50,102	1,638	5,626	0.0046	0.0140	0.1417	7	8
Unambiguous names	29,855	Not computed	Not computed	Not computed	Not computed	Not computed	7	8

The vocabulary is small: about 1,600 question tokens and 5,600 answer tokens. Most of the answer vocabulary is patient names and hospital or doctor names. The very low TTR reflects the fixed template wording, not a lack of content.

6.3 Behaviour diversity

Level	Finding
Metadata behavior	One label for all rows
Metadata question_type	One label for all rows
Question templates	14
Actual question intent	One intent per column: a value lookup (age, billing amount), a category lookup (gender, blood type), or a date or text lookup

The metadata shows one behaviour and one question type. The data has 14 different intents, one per column. Use column_focus to separate them.

6.4 Question pattern diversity

Pattern	Rows
One of 14 fixed templates	50,102 (all rows)
Questions that do not match one of the 14 templates after name replacement	0
Exact duplicate questions	0

6.5 Answer diversity

Answers are unique per row (50,102 unique answers), but they are not diverse in wording. Each answer is a template filled with a name and a value. Diversity comes from the values, such as ages, room numbers, dates, and billing amounts.

6.6 Overall rating (qualitative)

Aspect	Rating	Reason
Question wording variety	Very low	14 fixed templates
Value variety	High	Ages, dates, room numbers, billing amounts span wide ranges
Entity variety	High, but ambiguous	38,163 names, but 40.4% of rows share a name
Column coverage	Balanced	About 3,570 rows per column
Category balance	Balanced	Categories split close to evenly (Section 4.1)

---

7. Known Issues and Limitations

#	Issue	Impact	Suggested handling
1	Ambiguous patient names (20,247 rows, 40.4%). A name in a question may refer to several records.	Answers may be wrong for the intended record. Evaluation cannot be scored fairly.	Use load_unambiguous() (29,855 rows). See Section 2.
2	Name variants are not normalized. 991 names start with a title (mr., mrs., ms., dr.), and 1,049 include a suffix such as jr. or md. The same person may appear as jamie campbell and mrs. jamie campbell.	Two rows for one person may look like two patients. Name-based splits may not separate the same person.	Normalize titles and suffixes before grouping. Use source_name plus name as a grouping key, not the name alone.
3	Billing amounts described as approximate. All 3,572 billing answers say lagabhag (approximately) before a value with two decimal places.	The value is stated as exact in the source. The wording is misleading.	Treat the number as the value. Drop the word lagabhag in training data if the model should state exact values.
4	Medication names are phonetic romanizations (espirina, aaibuprophena, lipitar, penisilina, pyarasitamol).	These will not match drug databases or English medical text by exact string.	Map to standard names before any medical lookup.
5	Metadata does not match the text. script (Deva), content_script (देवनागरी), and behavior_definition describe Devanagari, but the text is romanized.	Filtering on these fields returns the wrong result.	Filter on column_focus or the text, not on script.
6	Dropped source positions are not explained. 1,698 of the 51,800 positions that could produce a row are absent.	The selection may not be random, and the coverage of the source is unknown.	Describe the subset as final_unique from the source, and do not claim full coverage.
7	Very low lexical diversity. 14 templates, and TTR (questions) of 0.0046.	A model may learn the templates, not the content.	Do not use this file alone to test language understanding. Use it for structured-field recall.
8	Synthetic or sensitive data. The records have names and clinical fields. The Kaggle source is generally described as synthetic, but this is not stated in the file.	Risk of treating synthetic data as real patient data, or of PII exposure if the data is real.	Confirm the origin on the Kaggle page before any use. Do not add real patient data.
9	Not medical advice. Medical conditions and medications appear in the answers.	Risk if used for clinical decisions.	Use for research, education, or model testing only.
10	Licence field. The file says Apache-2.0, but the source is a Kaggle dataset with its own terms.	The Apache-2.0 label does not confirm the source's terms.	Check the Kaggle page terms (Section 10).
11	Dates are not checked for logical order. Admission and discharge dates are separate rows, so they cannot be compared within one record here.	Some discharge dates could fall before admission dates, and the file does not show this.	If needed, rebuild the pairs from the source records.

---

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/healthcare_romanized_qa_final_unique.jsonl"

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

print(len(rows))                                   # 50102
print(rows[0]["conversations"][0]["value"])        # question
print(rows[0]["conversations"][1]["value"])        # answer
meta = json.loads(rows[0]["metadata_json"])        # metadata is a JSON string
print(meta["column_focus"])                        # Age

8.2 Using the included loader

load_healthcare_qa_romanized.py (included with this README) extracts the patient name, flags ambiguous names, and can keep only unambiguous rows.

from load_healthcare_qa_romanized import load_all, load_unambiguous, summarize

F = "/path/to/healthcare_romanized_qa_final_unique.jsonl"

records = load_all(F)
summarize(records)
print(sum(r["ambiguous_name"] for r in records))    # 20247

clean = load_unambiguous(F)
print(len(clean))                                   # 29855
print(clean[0]["instruction"], "->", clean[0]["output"])

Command line:

python load_healthcare_qa_romanized.py --file /path/to/healthcare_romanized_qa_final_unique.jsonl
python load_healthcare_qa_romanized.py --file /path/to/healthcare_romanized_qa_final_unique.jsonl --unambiguous-only

Expected output of the first command (first lines):

Total records: 50102
Ambiguous-name rows: 20247
Per column:
   3593  Age
   3590  Test Results
   3589  Hospital
   3588  Discharge Date
   3584  Gender
   3584  Medical Condition
   3578  Medication
   3574  Blood Type
   3574  Insurance Provider
   3572  Billing Amount
   3571  Date of Admission
   3571  Room Number
   3568  Admission Type
   3566  Doctor

8.3 HuggingFace datasets

import re
from collections import Counter
from datasets import load_dataset

TAILS = [r" ko umera", r" ko linga", r" ko rakta samuha", r" lai kuna rog",
         r" kahile aspatal", r" ko upachar garne", r" kuna aspatalma",
         r" ko bima pradayak", r" ko biling rakam", r" kuna kotha",
         r" ko bharna prakar", r" kahile discharja", r" lai kuna aushadhi",
         r" ko parikshan natija"]
TAIL_RE = re.compile("(" + "|".join(TAILS) + ")")

def name_of(question):
    m = TAIL_RE.search(question)
    return question[: m.start()] if m else None

ds = load_dataset("json", data_files={"hc": "/path/to/healthcare_romanized_qa_final_unique.jsonl"}, split="hc")
names = [name_of(r["conversations"][0]["value"]) for r in ds]
counts = Counter(names)
ds = ds.add_column("patient_name", names)
ds = ds.add_column("ambiguous_name", [counts[n] > 1 for n in names])
print(Counter(ds["ambiguous_name"]))     # False: 29855, True: 20247

8.4 Keep unambiguous rows and convert to instruction format

unambiguous = ds.filter(lambda r: not r["ambiguous_name"])

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

alpaca = unambiguous.map(to_alpaca, remove_columns=unambiguous.column_names)
print(alpaca[0])

8.5 Split by patient name (leak-safe)

Rows that share a name stay in the same split, so the same patient cannot appear in both train and test.

import random

all_names = sorted(set(n for n in names if n))
rng = random.Random(42)
rng.shuffle(all_names)
test_names = set(all_names[: int(0.1 * len(all_names))])

test = ds.filter(lambda r: r["patient_name"] in test_names)
train = ds.filter(lambda r: r["patient_name"] not in test_names)
print(train.num_rows, test.num_rows)     # 45110 4992

8.6 Filter by column

import json

billing = ds.filter(lambda r: json.loads(r["metadata_json"])["column_focus"] == "Billing Amount")
print(billing.num_rows)                  # 3572

8.7 Remove exact duplicate pairs

This dataset has no exact duplicates, but the check is cheap and safe to keep in a pipeline.

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

print(len(keep))                         # 50102

8.8 Pandas (first 1,000 rows)

import json
import pandas as pd

df = pd.read_json("/path/to/healthcare_romanized_qa_final_unique.jsonl", lines=True, nrows=1000)
df["question"] = df["conversations"].apply(lambda c: c[0]["value"])
df["answer"] = df["conversations"].apply(lambda c: c[1]["value"])
df["column_focus"] = df["metadata_json"].apply(lambda s: json.loads(s)["column_focus"])
print(df[["column_focus", "question", "answer"]].head())

---

9. Recommended Usage

Goal	Recommendation
Train a model to answer structured questions about records	Use the 29,855 unambiguous rows. Balance by column_focus.
Evaluate a model	Use name-based split (Section 8.5) on unambiguous rows. Report per column.
Test whether a model reads the right field	Use the 14 columns as separate tasks.
Test language understanding	Do not use this file. The 14 templates do not test language.
Medical use	Not suitable.

General advice:

• Report per column. Column difficulty and answer types differ.
• Do not evaluate on ambiguous rows.
• Confirm the synthetic or real origin of the source before any release.

---

10. Licensing, Privacy, and Sources

Item	Detail
Source	Kaggle, prasad22/healthcare-dataset: https://www.kaggle.com/datasets/prasad22/healthcare-dataset
License field in file	Apache-2.0
Where to check terms	The Kaggle dataset page, under its license and usage notes
Personal data	The file contains patient names and clinical fields. Confirm on the Kaggle page whether the records are synthetic. Do not add real patient data.

The license field describes the QA pairs as provided. It does not replace the terms of the Kaggle source. Check those terms before redistributing the file or publishing a model trained on it.

---

11. Reproducing the Statistics

This script uses the tokenizer in Section 6.1 and the name extraction in Section 8.3.

import json, re, statistics
from collections import Counter

PUNCT = re.compile(r"[?.,!:;()'\"/\-–]+")
TAILS = [r" ko umera", r" ko linga", r" ko rakta samuha", r" lai kuna rog",
         r" kahile aspatal", r" ko upachar garne", r" kuna aspatalma",
         r" ko bima pradayak", r" ko biling rakam", r" kuna kotha",
         r" ko bharna prakar", r" kahile discharja", r" lai kuna aushadhi",
         r" ko parikshan natija"]
TAIL_RE = re.compile("(" + "|".join(TAILS) + ")")

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

def name_of(q):
    m = TAIL_RE.search(q)
    return q[: m.start()] if m else None

rows = [json.loads(l) for l in open("healthcare_romanized_qa_final_unique.jsonl", encoding="utf-8") if l.strip()]
Q = [r["conversations"][0]["value"] for r in rows]
A = [r["conversations"][1]["value"] for r in rows]
names = [name_of(q) for q in Q]
counts = Counter(names)

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"rows={len(rows)} distinct_names={len(counts)} "
      f"ambiguous={sum(1 for n in names if counts[n] > 1)}")
print(f"qvocab={len(set(qt))} avocab={len(set(at))} "
      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}")
print(f"medQ={statistics.median(len(q.split()) for q in Q)} "
      f"medA={statistics.median(len(a.split()) for a in A)}")

Expected output: rows=50102 distinct_names=38163 ambiguous=20247, qvocab=1638 avocab=5626 TTR_q=0.0046 TTR_a=0.0140 bigram_q=0.1417, and medQ=7 medA=8.

---

12. Changelog

• v1 (current): Initial README for healthcare_romanized_qa_final_unique.jsonl. Identified the patient-name ambiguity (20,247 rows), the 14 fixed templates, the approximate wording used for exact billing values, the phonetic medication names, and the metadata script mismatch. Added load_healthcare_qa_romanized.py with an option to keep only unambiguous rows.