sabin1234/Pharmacy_Licensing_Exam_FAQ

💊 Pharmacy Licensing Exam FAQ (Nepali) — Dataset README

A 100% Nepali-language synthetic reasoning dataset built from a single row-pair of official Nepal Department of Health Services exam-result data — turned into 100 question-answer pairs covering counts, percentages, ratios, statistics, and "what-if" arithmetic.

---

🔖 TL;DR (At a Glance)

What	Answer
Total records	100
File size	~198 KB
Language / script	Nepali (ne / npi) — Devanagari
Format	JSONL, conversations (ShareGPT-style, 2-turn)
Task type	instruction-following
License	Creative Commons Attribution (permissive)
Data source	Department of Health Services — pharmacy licensing exam results
Generation type	synthetic (not real, unlike the NEPSE dividend dataset)
Source rows used	Only 2 rows (row_0, row_1) from one merged CSV
Fiscal year represented	2071/072 only
Distinct question "shapes"	Highly varied — 20+ reasoning patterns, not a fixed template set

---

🗂️ Schema — What's Inside Each Record

{
  "id": "sg_b39dea595f5f3f189ce06472bec1577e",
  "conversations": [
    {"from": "human", "value": "How many pharmacists attended the exam?"},
    {"from": "gpt",   "value": "110 pharmacists attended the exam."}
  ],
  "source": "Department of Health Services",
  "source_name": "pharmacy_licensing_exam_faq",
  "source_repo": "Department of Health Services",
  "source_config": "default",
  "source_split": "train",
  "source_revision": "v1",
  "source_row_id": "Licensing_Examination_Results_Pharmacists_Pharmacy_FY2071_72_to_2073_74_merged.csv:row_0,row_1:1",
  "language": "ne", "language_code": "npi", "script": "Deva",
  "license": "Creative Commons Attribution", "license_tier": "permissive",
  "task_type": "instruction-following",
  "generation_type": "synthetic", "condition": "synthetic",
  "url": "",
  "metadata_json": "{...}"
}

Field-by-field status

Field	Type	Variation	Value
id	string	✅ unique per record	hash-style, e.g. sg_b39dea595f5f3f189ce06472bec1577e (not sequential)
conversations	list[2]	✅ content differs, structure fixed	always human → gpt
source, source_repo	string	❌ constant	"Department of Health Services"
source_name	string	❌ constant	"pharmacy_licensing_exam_faq"
source_config / source_split / source_revision	string	❌ constant	default / train / v1
source_row_id	string	✅ varies (counter only)	same CSV + same 2 rows every time — only the trailing index (1→100) changes
language / language_code / script	string	❌ constant	ne / npi / Deva
license / license_tier	string	❌ constant	Creative Commons Attribution / permissive
task_type	string	❌ constant	instruction-following
generation_type / condition	string	❌ constant	synthetic
url	string	❌ empty	always ""
metadata_json	string (nested JSON)	❌ constant internally	see below

Inside metadata_json (nested) — also all constant

{
  "generation_domain": "Public Health Exam Statistics",
  "generation_category": "Pharmacy Licensing Exam Results",
  "generation_sub_domain": "Direct Q&A and basic statistical reasoning",
  "behavior": "Giving direct and natural answers",
  "behavior_definition": "Answering directly, naturally, and concisely in pure Nepali, using only the given values.",
  "question_type": "Direct conversational question",
  "content_language": "Nepali",
  "content_script": "Devanagari",
  "english_content_allowed": false
}

The generation_sub_domain — "basic statistical reasoning" — is the key differentiator from the NEPSE dividend dataset: this collection isn't pure fact-lookup, it's arithmetic/statistical reasoning over two numbers.

---

⚠️ The Most Important Finding: All 100 Records Come From Just 2 Source Rows

Unlike a typical FAQ dataset covering many companies/years, every single one of the 100 records is derived from exactly one row-pair in one CSV file:

Licensing_Examination_Results_Pharmacists_Pharmacy_FY2071_72_to_2073_74_merged.csv
  → row_0 (Pharmacist results)
  → row_1 (Pharmacy-candidate results)

Despite the filename suggesting the CSV merges FY 2071/072 through FY 2073/074, only the fiscal-year-2071/072 row pair was actually sampled — confirmed by record #9, whose answer states both groups belong to FY 2071/072. So this is not a broad exam-history dataset; it's a single exam sitting, exhaustively interrogated 100 different ways.

The underlying two data points

Group	Attended	Passed	Failed	Column value
Pharmacist	110	96	14	87
Pharmacy candidate	481	426	55	89

Everything else in the dataset — sums, differences, ratios, percentages, means, medians, hypotheticals — is arithmetic derived from just these figures.

---

💬 Conversation Structure

• Every record has exactly 2 turns: human → gpt. 100/100 consistent, no system prompt.
• Questions and answers are noticeably shorter and more varied in shape than a fixed-template dataset:

	Question	Answer
Word count (avg)	~8.0 words	~6.0 words
Character length (avg)	~46.6 characters	~34.4 characters
Min–max (chars)	17–80	10–123

• Unique questions: 100 / 100 (no exact duplicate questions)
• Unique answers: 93 / 100 — 7 answers repeat across unrelated questions (see Data Quality below)

---

❓ What Kinds of Questions Does It Ask?

Unlike the NEPSE dividend dataset (6 fixed templates), this dataset has no small fixed template set — it's built around a wide range of statistical/arithmetic reasoning operations applied to the same two numbers. Keyword frequency across the 100 questions:

Reasoning type (keyword)	Appears in questions
percentage (प्रतिशत)	12
rate (दर)	11
ratio (अनुपात)	10
difference (भिन्नता)	4
average/mean (औसत)	4
median (मध्यक)	3
range (दायरा)	3
product (गुणनफल)	3
sum (योग)	1

Beyond these, the dataset includes several other question families:

• Direct counts — "How many pharmacists attended/passed/failed?"
• Comparisons — "Which group had more/fewer attendees or passes?"
• Combined totals — "How many candidates attended/passed across both groups combined?"
• Pass-to-attendance ratios — expressed both as fractions and as X:Y ratios
• Percentage-point gaps — "How many percentage points higher is the pharmacy pass rate?"
• Hypothetical / what-if scenarios — "If 5 more pharmacists had passed, what would the new pass rate be?", "If both groups improved to a 90% pass rate, how many total passes would that require?"
• Statistics — mean, median, and range across the two groups' figures
• Products — "What is the product of the two attendance figures?" (110 × 481 = 52,910)

This is a reasoning-over-two-numbers dataset, not a fact-recall dataset — the opposite design philosophy from the NEPSE dividend FAQ collection.

---

🔁 Duplicates & Data Quality

Check	Result
Unique id	100 / 100 ✅
Unique questions	100 / 100 ✅
Unique answers	93 / 100 ⚠️
Source data points	Only 2 rows (1 exam sitting)

⚠️ Mismatched/copy-pasted answers (7 duplicate answers, 15 affected questions)

The most notable quality issue: several distinct questions share an identical answer that doesn't actually match what was asked. For example:

Answer (repeated 6×)	Was given for these questions
"The pharmacy-candidate group had more attendees."	"Which group had more attendees?" ✅ correct — but also given verbatim for "Which group had more passes?", "Which group had a higher pass rate?", "Which group had a lower fail rate?", "What share of all attendees were pharmacy candidates?", and "What percentage of pharmacy candidates failed?" ❌

Two smaller cases follow the same pattern — an answer about "column value" or "fewer attendees" gets reused for a differently-worded question about passes. This looks like a generation-pipeline bug where the answer for one templated question in a sequence occasionally leaked into a neighboring question — 7 of the 100 answers (affecting 15 questions total) are affected.

⚠️ Garbled/corrupted tokens despite english_content_allowed: false

8 records contain stray, non-Nepali fragments embedded mid-sentence — most likely corrupted renderings of English connector words like "more" or "or":

"४१ mवाe फार्मेसी उम्मेदवार पास भएनन्।"
"फार्मासिस्ट ओगटेका fवा लगभग २०.२९% मध्ये नपास उम्मेदवार।"
"२७५ mवाe फार्मेसी उपस्थित।"

These fragments (mवाe, fवा) break the dataset's stated Devanagari-only guarantee and read as broken/untranslated artifacts rather than intentional content — worth cleaning or filtering before downstream use.

Other notes

• 92 / 100 answers contain at least one number (Devanagari or Arabic numerals) — a handful of qualitative/comparison answers have none.
• No missing fields, no turn-structure violations, clean UTF-8 elsewhere.

---

🎨 Design Philosophy (the likely thinking behind this dataset)

1. Depth over breadth — instead of covering many exams across many years (like the NEPSE dataset covers many companies), this dataset takes one exam result and asks it 100 different ways, testing whether a model can perform varied arithmetic/statistical reasoning over the same two numbers.
2. Reasoning-focused, not recall-focused — heavy use of ratios, percentages, means, medians, ranges, and hypothetical "what if N more had passed" scenarios, rather than simple fact lookup.
3. Real-world grounding, synthetic construction — the underlying numbers come from a real Department of Health Services exam-results file, but the Q&A pairs themselves are machine-generated (generation_type: synthetic), which likely explains the answer-mismatch and garbled-token issues found above.
4. Nepali-only intent, imperfectly executed — english_content_allowed: false is declared, but a small number of records leak corrupted non-Nepali fragments, suggesting a translation/generation step didn't fully complete.

---

🧪 Sample Records

{"id": "sg_b39dea595f5f3f189ce06472bec1577e",
 "conversations": [
   {"from": "human", "value": "How many pharmacists attended the exam?"},
   {"from": "gpt", "value": "110 pharmacists attended the exam."}
 ]}

{"id": "sg_...", 
 "conversations": [
   {"from": "human", "value": "What is the ratio of pharmacy to pharmacist attendance?"},
   {"from": "gpt", "value": "The attendance ratio is 481:110, or approximately 4.37:1."}
 ]}

---

⚠️ Things to Watch Out For When Using This Dataset

• Extremely narrow source coverage — all 100 Q&A pairs derive from a single exam sitting (2 rows). This is not representative of pharmacy licensing results across years or exam types; treat it as a reasoning-skill dataset, not a knowledge-coverage dataset.
• ~7% of answers are mismatched to their question due to an apparent generation bug — worth filtering or manually correcting before training/production use.
• 8% of records contain corrupted non-Nepali fragments (mवाe, fवा) — worth a cleanup pass.
• Synthetic, not real — unlike a real-world-grounded FAQ set, some hypothetical/statistical answers (e.g., "if both groups improved to 90%...") involve derived/approximated figures rather than directly-reported facts; verify precision-sensitive values before relying on them.

---

📜 License

Creative Commons Attribution (permissive) — free to use with attribution.

---

This README was compiled from a full programmatic analysis of all 100 records in the dataset (all figures calculated directly from the file, not estimated).