sabin1234/NEPSE_Dividend_FAQ

📈 NEPSE Dividend FAQ (Nepali) — Dataset README

"How much dividend did it give this time?" — the question every stock market watcher asks, now turned into 2,000 question-answer pairs.

A simple, clean, 100% Nepali-language financial FAQ dataset that builds question-answer (Q&A) pairs about the total dividend and bonus share dividend of companies listed on NEPSE. This file is a full dissection of that exact dataset — every field, every pattern, everything.

---

🔖 TL;DR (At a Glance)

What	Answer
Total records	2,000
File size	~3.27 MB
Language / script	Nepali (ne / npi) — Devanagari
Format	JSONL, conversations (ShareGPT-style, 2-turn)
Task type	instruction-following
License	Apache-2.0 (permissive)
Data source	NEPSE Dividend History (real-world data — generation_type: real)
Companies covered	211 distinct company codes
Fiscal year range covered	FY 2062/063 to FY 2082/083
Question templates	Exactly 6
Answer templates	Exactly 2

---

🗂️ Schema — What's Inside Each Record

Each line is an independent JSON object, and all of them share the same structure:

{
  "id": "nepse_dividend_0001",
  "conversations": [
    {"from": "human", "value": "What was ACLBSL's total dividend for fiscal year 2077/078?"},
    {"from": "gpt",   "value": "ACLBSL's total dividend for fiscal year 2077/078 was 21%."}
  ],
  "source": "NEPSE Dividend History",
  "language": "ne", "language_code": "npi", "script": "Deva",
  "license": "Apache-2.0", "license_tier": "permissive",
  "task_type": "instruction-following",
  "generation_type": "real", "condition": "real",
  "metadata_json": "{...}"
}

Field-by-field status

Field	Type	Variation	Value
id	string	✅ unique per record	nepse_dividend_0001 → nepse_dividend_2000 (sequential)
conversations	list[2]	✅ content differs, structure fixed	always human → gpt, never more/fewer turns
source, source_name, source_repo	string	❌ constant	"NEPSE Dividend History" / "nepse_dividend_faq"
source_config, source_split, source_revision	string	❌ constant	default, train, v1
language / language_code / script	string	❌ constant	ne / npi / Deva
license / license_tier	string	❌ constant	Apache-2.0 / permissive
task_type	string	❌ constant	instruction-following
generation_type / condition	string	❌ constant	real (not synthetic — based on real-world data)
url	string	❌ empty	always ""
metadata_json	string (nested JSON)	❌ constant internally	see below

Interesting fact: almost all dataset-metadata fields (source, license, language, task_type, generation_type...) are 100% constant — meaning this is a homogeneous dataset built from a single "batch" using a single recipe. All the variation lives in the actual Q&A content — the company, the fiscal year, and the dividend percentage.

Inside metadata_json (nested) — also all constant!

{
  "generation_domain": "Financial Services",
  "generation_category": "NEPSE Dividend History",
  "generation_sub_domain": "Company Dividends",
  "behavior": "Providing factual dividend information",
  "behavior_definition": "Directly answering the question using dividend data for the specified company and fiscal year.",
  "question_type": "Fact-based question",
  "content_language": "Nepali",
  "content_script": "Devanagari",
  "english_content_allowed": false
}

All 9 of these sub-fields are identical across 2,000 out of 2,000 lines — the behavior is a single, focused capability: "providing fact-based company dividend information." There's no multi-behavior mixing, and no opinion/analysis-style questions — pure facts only.

---

💬 Conversation Structure

• Each record has exactly 2 turns: human (question) → gpt (answer). Never 3-turn, 1-turn, or follow-up — this pattern holds for 100% of 2,000 records.
• There is no system prompt.
• Length is also very tight/consistent:

	Question	Answer
Word count (avg)	~11.6 words	~10.3 words
Character length (avg)	~67 characters	~60 characters
Min–max (chars)	54–85	53–70

Short, direct, single-fact Q&A — this is not a long conversation or reasoning-chain dataset. It's a quick-recall FAQ style.

---

❓ Question Pattern Distribution — Exactly 6 Templates

Once you strip out the company code and fiscal year, all 2,000 questions fit into one of these 6 templates (100% coverage, no outliers):

#	Template	Count	%
1	What was <COMPANY>'s total dividend for fiscal year <FY>?	468	23.4%
2	What total dividend percentage did <COMPANY> declare for fiscal year <FY>?	448	22.4%
3	What is <COMPANY>'s total dividend percentage for fiscal year <FY>?	401	20.1%
4	What is <COMPANY>'s bonus share dividend percentage for fiscal year <FY>?	231	11.6%
5	What was <COMPANY>'s bonus share dividend for fiscal year <FY>?	231	11.6%
6	What bonus share percentage did <COMPANY> declare for fiscal year <FY>?	221	11.1%

Template 1 ██████████████████████ 23.4%
Template 2 █████████████████████ 22.4%
Template 3 ████████████████████ 20.1%
Template 4 ███████████ 11.6%
Template 5 ███████████ 11.6%
Template 6 ███████████ 11.1%

Templates 1–3 = 3 phrasings asking about Total Dividend (1,317 records — 65.9%) Templates 4–6 = 3 phrasings asking about Bonus Share dividend (683 records — 34.1%)

For each "content" type (total/bonus), exactly 3 syntactic variations are provided — to give sentence-level diversity. But semantically, they're all just two kinds of questions.

---

✅ Answer Pattern Distribution — Exactly 2 Templates

The answers are even cleaner — there are only 2 templates total:

Template	Count	%
<COMPANY>'s total dividend for fiscal year <FY> was <PCT>.	1,317	65.9%
<COMPANY>'s bonus share dividend for fiscal year <FY> was <PCT>.	683	34.1%

No hedging, no extra context, no "this is an estimate"-style disclaimers — a straight fact → answer, in a deterministic format. This design reveals the dataset's purpose: "given a specific company + specific year, the model should give a clean, single-line, fact-based answer."

---

🏢 Company Coverage

• 211 distinct company codes are covered (e.g., EBL, NABIL, ADBL, GBIME...).
• Each company has an average of ~9.5 questions, but the distribution is quite skewed — some companies have data spanning many years, others have only one or two.

Top 15 Companies by Question Count

Company Code	Record Count
EBL	32
GBBL	31
CBBL	30
CHCL	30
CIT	30
DDBL	30
GBIME	30
ICFC	30
MDB	30
NABIL	30
CZBIL	29
BPCL	28
FMDBL	28
AHPC	26
HDL	26

Records-per-Company Density (histogram)

Records per Company	Number of Companies
1	20
2	30
3	20
4	16
5–9	45
10–19	44
20–29	15
30–32	21

Long-tail distribution — 20 companies have only 1 question each, while top-tier companies (large banks/finance companies like EBL, NABIL, GBIME) have up to 30 years of history covered. This reflects real-world data availability — older/larger companies have more historical records available, while newer/smaller ones have less.

---

📅 Fiscal Year Coverage

Fiscal years from 2062/063 to 2082/083 (nearly a 20-year range) are covered — but density is higher toward recent years:

Fiscal Year	Records
2062/063 – 2066/067	3
2067/068	67
2068/069	75
2069/070	74
2070/071	92
2071/072	99
2072/073	116
2073/074	129
2074/075	131
2075/076	149
2076/077	160
2077/078	178
2078/079	180
2079/080	134
2080/081	172
2081/082	206 (highest!)
2082/083	35 (most recent, still thin)

2067–069  ████████████████
2070–072  ███████████████████████
2073–075  ███████████████████████████████
2076–078  ██████████████████████████████████████
2079–081  ████████████████████████████████████████
2082–083  ███████

FY 2081/082 has the most records (206), while the earliest years (062-067) have very few (just 3 total) — a natural recency bias, since older data is less digitized/available.

---

💰 Dividend % Distribution

Bucket	Records	%
0% (no dividend)	114	5.7%
0.01% – 5%	181	9.0%
5.01% – 10%	446	22.3%
10.01% – 15%	458	22.9% ⭐ densest
15.01% – 20%	279	14.0%
20.01% – 30%	290	14.5%
30%+	232	11.6%

• Median: 13.5% | Average: ~19.7% (pulled upward by outliers)
• Most frequent single values: 10% (124 times), 15% (120 times), 0% (114 times), 20% (102 times)
• 392 distinct % values appear in total — many of them include decimals (e.g., 10.53%, 21.05%, 15.79%), which look like exact figures from actual bonus-share conversions, not rounded numbers.

🥇 The Highest Dividend (an interesting outlier!)

UNL appears to have declared a total dividend of 990% in FY 2071/072 and 860% in FY 2070/071.

This could be a data-entry error (e.g., 99.0% mistakenly entered as 990%), or it could genuinely reflect a bonus-heavy year for a micro-finance-type company back then. If you plan to use this dataset for training/production, it's a good idea to audit such extreme outliers.

---

🎁 Bonus Share vs. Total Dividend

Type	Records	%
Total Dividend	1,317	65.9%
Bonus Share Dividend	683	34.1%

• Only 1 total-dividend record shows 0% — almost every company/year has some form of dividend.
• But 113 bonus-share records show 0% — meaning many companies didn't distribute bonus shares in certain years (cash dividend only), which matches real NEPSE practice.

---

🔁 Duplicates & Data Quality

Check	Result
Unique id	2,000 / 2,000 ✅ (no duplicates)
Unique questions	1,998 / 2,000 (2 questions are exact duplicates)
Unique answers	1,999 / 2,000
Empty/missing fields	None found
Turn structure violations	None (always human→gpt)
Encoding	Clean UTF-8 Devanagari, no mixed-script/garbage characters

Overall, this dataset is very clean — almost production-ready for building FAQ pairs. Aside from the ~990% outlier mentioned above, no major errors were found.

---

🎨 Design Philosophy (the likely thinking behind this dataset)

1. Single, focused behavior — it does one thing: given a company + year, return the % dividend. No multi-task mixing.
2. Controlled syntactic diversity, fixed semantics — 6 question templates but the same underlying meaning (3 total + 3 bonus) — this seems aimed at making the model phrasing-robust (able to answer correctly no matter how the same information is asked).
3. Real-world grounding — generation_type: real — not synthetic/fictional data, but drawn from actual NEPSE dividend history (hence the outliers/uneven distribution — the real world is messy!).
4. Deterministic, hedge-free answers — no words like "approximately" or "estimated." This establishes raw ground truth for factual-recall tasks.
5. Nepali-only, Devanagari-only — english_content_allowed: false — no Roman/English mixed in anywhere, full linguistic purity in Nepali is maintained.

---

🧪 Sample Records

{"id": "nepse_dividend_0001",
 "conversations": [
   {"from": "human", "value": "What was ACLBSL's total dividend for fiscal year 2077/078?"},
   {"from": "gpt", "value": "ACLBSL's total dividend for fiscal year 2077/078 was 21%."}
 ]}

{"id": "nepse_dividend_0231",
 "conversations": [
   {"from": "human", "value": "What is NICL's bonus share dividend percentage for fiscal year 2074/075?"},
   {"from": "gpt", "value": "NICL's bonus share dividend for fiscal year 2074/075 was 10.53%."}
 ]}

---

⚠️ Things to Watch Out For When Using This Dataset

• Coverage is uneven — some companies/years have a lot of data, others just one record. Watch out for class imbalance when fine-tuning.
• Only company codes are given, not full names (e.g., "EBL" — presumably Everest Bank Limited) — if full names are needed, a separate mapping table is required.
• Outliers like the 990% figure should be audited.
• 2-turn, single-fact QA only — this dataset doesn't teach multi-turn conversation, reasoning, or comparative analysis ("which company paid more?"). If that kind of capability is needed, it should be combined with another dataset.

---

📜 License

Apache-2.0 (permissive) — free for both commercial and non-commercial use, with attribution.

---

This README was auto-generated — based on a programmatic analysis of the full 2,000 records (all figures calculated from the actual dataset, not estimated).