sabin1234/Healthy_Choice_Aesthetic_Hospital_FAQ

Healthy Choice Aesthetic Hospital FAQ — Pure Nepali Q&A Dataset

1. Overview

This dataset is a pure Nepali-language (Devanagari script) collection of question–answer pairs in ShareGPT format, covering frequently asked questions about Healthy Choice Aesthetic Hospital — a real aesthetic/cosmetic and general specialty hospital located in Kathmandu, Nepal. Topics span hospital directions/location, hair transplant, skin treatments, plastic surgery, laser procedures, dental, cardiology, and other clinical services offered by the hospital.

Each record is a single-turn conversation where a human asks a practical, patient-facing question in Nepali, and the assistant ("gpt") gives a concise, fact-based answer in Nepali.

Property	Value
File name	healthy_choice_faq_suddha_nepali.jsonl
File format	JSON Lines (.jsonl) — one JSON object per line
Total records (rows)	101
Conversation format	ShareGPT (conversations list with from/value pairs)
Turns per conversation	2 (1 human turn + 1 gpt turn) — 100% of records
Domain	Health — Aesthetic/cosmetic hospital FAQ (Healthy Choice Aesthetic Hospital, Kathmandu)
Content language	Nepali (Devanagari script) — 100% verified
Underlying data source	Cohere Labs' Aya Dataset (CohereLabs/aya_dataset), Nepali subset
License (per-record)	Apache-2.0 (permissive)
Duplicate IDs found	0 (all 101 id values unique)
Duplicate questions found	0
Malformed / unparseable lines	0

---

2. File & Record Structure

This dataset uses the same flat schema style seen in the other three "sourced/real-content" FAQ datasets reviewed previously (Fight Vitiligo, Global Healthcare Jobs, Harley Street Institute), with provenance metadata at the top level and fine-grained generation metadata nested in a stringified metadata_json field.

Key	Type	Description
id	string	Unique record identifier, ShareGPT-style hash format sg_<32-char hex> (e.g. sg_900a97b9e2420ec424d75ca45ea6aa76).
conversations	array	ShareGPT-style conversation turns. Always exactly 2 turns.
conversations[0].from	string	Always "human" — the question turn.
conversations[0].value	string	The Nepali-language question about the hospital, its location, or its services.
conversations[1].from	string	Always "gpt" — the answer turn.
conversations[1].value	string	A Nepali-language factual answer (length varies widely — see Section 3.3).
source	string	Origin dataset path.
source_name	string	Human-readable source name.
source_repo	string	Repository the source data was pulled from.
source_config	string	Config label.
source_split	string	Data split label.
source_revision	string	Fixed revision hash shared by all records in this file.
source_row_id	string	Row-level traceability ID, format <id>:1.
language	string	ISO language code.
language_code	string	ISO 639-3 code.
script	string	Script tag.
license	string	License of the content.
license_tier	string	License permissiveness bucket.
task_type	string	NLP task type this record represents.
generation_type	string	How the record was produced.
condition	string	Generation condition tag.
url	string	Source URL reference (always empty in this file).
metadata_json	string	Stringified JSON object with fine-grained generation metadata (parsed in Section 3).

2.1 Top-level provenance fields (all constant across the dataset)

Field	Observed value	%
source	CohereLabs/aya_dataset:default:train	100%
source_name	aya_human_nepali	100%
source_repo	CohereLabs/aya_dataset	100%
source_config	default	100%
source_split	train	100%
source_revision	f9ea04583f02a8f86404ff6c58bf75fe637df8a2	100%
language	ne	100%
language_code	npi	100%
script	Deva (Devanagari)	100%
license	Apache-2.0	100%
license_tier	permissive	100%
task_type	instruction-following	100%
generation_type	synthetic	100%
condition	synthetic	100%
url	"" (empty)	100%

Notable difference from the other three "sourced FAQ" datasets reviewed: this is the only one whose underlying source traces to a well-known public multilingual instruction dataset — Cohere Labs' Aya Dataset (aya_human_nepali subset), rather than a company/institute's own FAQ page. It is also the only one of the four tagged generation_type: synthetic / condition: synthetic — matching the pattern seen in the earlier BFI banking dataset rather than the "real"/"original" tags seen in Fight Vitiligo, Global Healthcare Jobs, and Harley Street Institute.

2.2 metadata_json sub-fields (parsed)

Field	Description	Observed value(s)
generation_domain	Top-level subject domain.	स्वास्थ्य ("Health") — 100%
generation_category	Sub-topic category.	एस्थेटिक अस्पताल प्रश्नोत्तर ("Aesthetic hospital Q&A") — 100%
generation_sub_domain	Specific entity/theme.	हेल्दी च्वाइस एस्थेटिक अस्पताल ("Healthy Choice Aesthetic Hospital") — 100%
behavior	Expected assistant behavior.	छोटो उत्तर माग्ने ("Request a short answer") — 100%
behavior_definition	Full definition of the expected behavior.	"मुख्य जानकारी कायम राख्दै संक्षिप्त उत्तर दिने।" ("Give a concise answer while retaining the key information.") — 100%
question_type	Format classification.	प्रश्नोत्तर प्रश्न ("Question-answer question") — 100%
question_length	Declared character-length bucket for the question.	५० देखि ३०० अक्षर ("50 to 300 characters") — 100%
response_length	Declared character-length bucket for the answer.	५० देखि ८०० अक्षर ("50 to 800 characters") — 100%
maximum_response_sentences	Max sentences allowed in the answer.	15 — 100%
minimum_reasoning_dimensions	Minimum reasoning dimensions required.	1 — 100%
minimum_complexity_score	Minimum complexity score assigned.	3 — 100%
content_language	Declared content language.	नेपाली ("Nepali") — 100%, accurate
content_script	Declared content script.	देवनागरी ("Devanagari") — 100%, accurate
english_content_allowed	Whether English content is permitted.	False — 100%

This dataset's generation constraints allow a notably wider and larger permitted answer length (up to 800 characters / 15 sentences) than the other three "real"/"original" FAQ datasets, and the actual maximum observed answer (1,223 characters — see Section 3.3) exceeds even that declared 800-character upper bound, indicating the length bucket is a soft/approximate target rather than a hard limit.

---

3. Question Distribution

3.1 By domain/category/sub-domain

The dataset is single-domain, single-entity, and single-topic by design — every record belongs to the same taxonomy path:

Level	Value	Count	%
generation_domain	स्वास्थ्य (Health)	101	100%
generation_category	एस्थेटिक अस्पताल प्रश्नोत्तर (Aesthetic hospital Q&A)	101	100%
generation_sub_domain	हेल्दी च्वाइस एस्थेटिक अस्पताल (Healthy Choice Aesthetic Hospital)	101	100%
question_type	प्रश्नोत्तर प्रश्न (Question-answer question)	101	100%

Unlike the earlier BFI banking MCQ dataset, there is no multiple-choice format here — every answer is free-text, and unlike Fight Vitiligo/Global Healthcare Jobs/Harley Street Institute, all questions are anchored to one single named business (a specific hospital) rather than a general topic area.

3.2 Thematic sub-topic breakdown (derived from question content)

The table below is a keyword-based approximate clustering of the actual question text (a question may match more than one theme, so counts can overlap):

Sub-theme (approx.)	Example keyword(s) matched	Questions matched	% of dataset
Surgery / procedures	शल्यक्रिया / अपरेसन	22	21.8%
Hair transplant	कपाल प्रत्यारोपण / कपाल	12	11.9%
Opening hours / timing	समय / खुल्छ / बन्द	7	6.9%
Treatment results / side effects	नतिजा / साइड इफेक्ट / असर	7	6.9%
Diet / nutrition	आहार / पोषण	6	5.9%
Dental	दन्त	5	5.0%
Cost / price	खर्च / शुल्क / मूल्य / लागत	4	4.0%
Skin treatment	छाला	4	4.0%
Department / services offered	विभाग / सेवा	3	3.0%
Doctor / specialist	डाक्टर / विशेषज्ञ / चिकित्सक	2	2.0%
Location / directions	कहाँ अवस्थित / नक्सा / बाटो / निर्देशन	1	1.0%
Appointment / booking	अपोइन्टमेन्ट / बुकिङ / भेटघाट	1	1.0%
Other / unmatched by keyword list	—	45	44.6%

56 of 101 questions (55.4%) matched at least one of the above keyword groups. Surgery/procedures and hair transplant are the dominant identifiable themes — reflecting the hospital's core aesthetic-medicine service lines (hair transplant methods like FUE/DHI, laser tattoo removal, non-surgical fat reduction/cryolipolysis, permanent hair removal, mole removal, acne and melasma treatment). Many "other/unmatched" questions cover further procedure-specific detail not captured by the simple keyword list (e.g. HydraFacial benefits, andrology consultations, donor/recipient area terminology in hair transplant, helmet-wearing safety timing after transplant).

3.3 Length characteristics

Metric	Min	Median	Mean	Max
Question — character length	18	46	46.7	92
Question — word count	3	7	7.5	14
Answer — character length	43	131	181.0	1,223
Answer — word count	6	20	26.6	185
Answer — sentence count (approx.)	1	1	1.9	18

Answer lengths are highly skewed — most answers are short (median 131 characters, 20 words), but a small number of records are much longer, dominated by one clear outlier: the hospital's directions/location answer (1,223 characters, turn-by-turn navigation instructions from two different Kathmandu landmarks — Naxal/Bhatbhateni and Chabahil). Excluding that single outlier, the next-longest answers (669, 623, 558, 514 characters) relate to detailed procedural explanations (e.g. andrologist consultation timing, hair transplant methods comparison). This wide variance (max 1,223 vs. declared bucket ceiling of 800 characters) shows the dataset mixes very short factual answers with occasional long, structured, multi-part explanations as needed by the question.

3.4 Answer format

Answers are free-form and practical, ranging from single-sentence direct facts (e.g. cost or session-count answers) to long, structured, multi-step explanations (e.g. the detailed floor-by-floor, landmark-by-landmark directions to the hospital, which reads almost like a step-by-step wayfinding guide). No lettered options or MCQ structure is present anywhere in the dataset.

---

4. Behavior Distribution

The dataset targets a single, uniform assistant behavior:

Behavior	Behavior definition	Count	%
छोटो उत्तर माग्ने ("Request a short answer")	"Give a concise answer while retaining the key information."	101	100%

4.1 Behavior-related generation constraints

Constraint	Value	Meaning
maximum_response_sentences	15	A relatively generous sentence cap — allows multi-step or multi-part answers (e.g. the long directions answer) while still bounding total length.
minimum_reasoning_dimensions	1	Each answer requires only 1 reasoning dimension — direct factual lookup/recall, not multi-fact combination.
minimum_complexity_score	3	A mid-low complexity tier — higher than the Fight Vitiligo dataset's score of 1, but well below the BFI banking dataset's score of 5.
question_length	50–300 characters (declared)	Actual observed range: 18–92 characters — narrower than the declared bucket.
response_length	50–800 characters (declared)	Actual observed range: 43–1,223 characters — exceeds the declared upper bound for the longest (directions) answer.
english_content_allowed	False	No English content permitted anywhere in the record — output must be pure Nepali.

4.2 Task type / generation type

Field	Value	Count	%
task_type	instruction-following	101	100%
generation_type	synthetic	101	100%
condition	synthetic	101	100%

Despite being labeled synthetic (as opposed to real/original in the other business-FAQ-style datasets reviewed), the content is clearly grounded in real, specific facts about an actual hospital (its physical layout, block/floor structure, nearby landmarks, and service offerings), suggesting the Q&A pairs were synthetically generated from real underlying hospital information rather than being freely invented.

---

5. Provenance & Licensing Distribution

Field	Value	Count	%
source	CohereLabs/aya_dataset:default:train	101	100%
source_name	aya_human_nepali	101	100%
source_repo	CohereLabs/aya_dataset	101	100%
source_split	train	101	100%
source_revision	f9ea04583f02a8f86404ff6c58bf75fe637df8a2	101	100%
license	Apache-2.0	101	100%
license_tier	permissive	101	100%

• Single source: all 101 records trace back to the same underlying source snapshot — the Nepali-language human-written subset of Cohere Labs' publicly available Aya Dataset, a large multilingual instruction-tuning dataset.
• License: Apache-2.0, matching the license tier used in the Fight Vitiligo, Global Healthcare Jobs, and Harley Street Institute datasets.
• Traceability: source_row_id mirrors the top-level id with a :1 suffix, and all records share one fixed source_revision hash — indicating this file is a single filtered/extracted batch (the Healthy Choice Aesthetic Hospital-related subset) pulled from the larger Aya Dataset Nepali collection.
• No URL field populated: every record's url field is an empty string.

---

6. Language & Script Notes

This dataset's language metadata is accurate and consistent with its actual content, matching the pattern seen in Fight Vitiligo, Global Healthcare Jobs, and Harley Street Institute:

Field	Declared value	Actual content observed
language	ne	Nepali ✅
language_code	npi	Nepali (ISO 639-3: npi) ✅
script	Deva	Devanagari ✅
metadata_json.content_language	नेपाली (Nepali)	Nepali ✅
metadata_json.content_script	देवनागरी (Devanagari)	Devanagari ✅
metadata_json.english_content_allowed	False	No English text found ✅

Verified finding: A full script scan of all 101 records confirms 100% of both questions and answers are written entirely in Devanagari script — zero Latin-only text in either turn. This matches the file name (suddha_nepali, i.e. "pure Nepali") accurately. As in the other real-content datasets, some medical/procedural terminology is transliterated phonetically into Devanagari rather than translated (e.g. "क्रायोलिपोलिसिस" for "cryolipolysis," "हाइड्राफेशियल" for "HydraFacial," "एन्ड्रोलोजिस्ट" for "andrologist," "एफयुई," "डीएचआई" for hair-transplant technique acronyms FUE/DHI) — a natural localization choice for specialized clinical terms, not a script violation.

---

7. Content Coverage Notes

• Single-institution focus: every question concerns one specific, named hospital — Healthy Choice Aesthetic Hospital in Kathmandu — including its physical location/directions, floor-by-floor department layout, and service offerings.
• Service lines covered: hair transplant (FUE, DHI, Instant DHT methods), skin treatments (melasma, acne, moles, HydraFacial), laser procedures (tattoo removal, permanent hair removal via Soprano), non-surgical body-fat reduction (cryolipolysis), plastic surgery, dental, cardiology, ophthalmology, and general andrology/consultation guidance.
• Detailed real-world grounding: the directions/location answer includes specific, verifiable local landmarks (Bhatbhateni supermarket, Nabil Bank, Chabahil, Om Hospital, "yellow bridge") and block/floor department mapping — indicating this content is grounded in the hospital's actual physical layout rather than being generic template text.
• No duplicate records or questions: all 101 id values and all 101 question texts are unique.
• No malformed JSON lines: every line parses successfully with a consistent flat schema.

---

8. Suggested Use Cases

• Fine-tuning or evaluating Nepali-language LLMs for local business/healthcare-facility customer-support chatbots (location/directions, service listings, pricing FAQs).
• Domain-adaptation for Nepali-language hospital or clinic virtual assistants, particularly for aesthetic/cosmetic and hair-restoration services.
• Benchmarking variable-length factual answer generation in Nepali — from single-fact short answers to long, structured, step-by-step directions.
• As a template for building similarly structured single-business FAQ datasets for other Nepali healthcare facilities, using the same Aya-Dataset-derived schema.
• As part of a combined Nepali health-FAQ corpus alongside Fight Vitiligo, Global Healthcare Jobs, and Harley Street Institute datasets, to study variation in answer style, length distribution, and generation-type labeling (synthetic vs. real/original) across differently-sourced Nepali health content.

---

9. Known Limitations

• Small dataset size: only 101 records — useful for narrow fine-tuning/evaluation on this specific hospital's FAQ, but too small alone for broad instruction tuning.
• Single domain, single entity, single behavior: entirely focused on one named hospital's FAQ with one behavior type (छोटो उत्तर माग्ने) — should be combined with other datasets for general-purpose or multi-domain instruction tuning.
• Highly skewed answer length: a small number of long answers (up to 1,223 characters) coexist with many very short ones (as few as 43 characters); models trained naively on this distribution may need length-aware sampling or weighting to avoid over-indexing on the outlier long-form directions answer.
• Declared length bucket exceeded: the metadata's stated response_length ceiling (800 characters) does not hold for at least one record (1,223 characters), so this field should be treated as an approximate target rather than a guaranteed hard bound.
• Highly location/time-specific content: directions, landmarks, and possibly pricing information reflect a specific point in time and physical layout; such details can become outdated if the hospital relocates, renovates, or changes pricing.
• No explicit source URLs: the url field is empty for every record, so individual FAQ entries cannot be traced back to a specific original web page.