sabin1234/NEPSE_Dividend_FAQ_Dataset_Romanized_Nepali_Questions

NEPSE Dividend FAQ Dataset (Romanized Nepali Questions)

File: nepse_dividend_faq_romanized.jsonl Total records: 2,000 Format: JSON Lines (.jsonl) — one JSON object per line Language: Nepali (ne / ISO 639-3 npi), answers in Devanagari script (Deva); questions in romanized Nepali (Latin letters) Domain: Financial services — NEPSE (Nepal Stock Exchange) historical dividend records Task type: Instruction-following (instruction-following) Generation type: Real (real) — this is factual data derived from actual company dividend history, not synthetically generated text License: Apache-2.0 (permissive)

This dataset is a single-turn (1 question + 1 answer) factual FAQ dataset built from real historical dividend announcements of companies listed on the Nepal Stock Exchange (NEPSE). Every question asks for the total dividend percentage (or bonus share percentage) that a specific listed company declared for a specific Nepali fiscal year (Bikram Sambat, e.g. 2077/2078), and every answer states the factual dividend percentage. A distinctive feature of this dataset is that questions are written in romanized Nepali (Latin script) while answers are written in standard Nepali (Devanagari script) — making it useful for romanized-input / Devanagari-output transliteration-aware QA systems.

---

1. Table of Contents

Section	Description
2. Dataset Summary	High-level counts and coverage
3. File / Record Schema	Full field list, types, descriptions
4. Nested metadata_json Schema	Fields inside the embedded metadata string
5. Conversation Structure	Turn-by-turn layout (single-turn Q&A)
6. Question Pattern Diversity	The 3 phrasing templates × 2 dividend types = 6 question forms
7. Answer Pattern Diversity	The fixed factual-answer template
8. Company Coverage	210 unique NEPSE-listed companies referenced
9. Fiscal Year Coverage	18 unique Nepali fiscal years covered
10. Dividend Value Coverage	Range and distribution of dividend percentages
11. Length & Size Statistics	Character-length statistics
12. Metadata Field Value Reference	Full list of fixed metadata values
13. Loading the Dataset (Code Demos)	Python, pandas, and Hugging Face datasets examples
14. Example Records	Fully annotated examples for both dividend types
15. Suggested Use Cases	Fine-tuning, evaluation, etc.
16. Limitations & Notes	Romanization ambiguity, no company-name mapping, etc.
17. Quick Reference Cheat Sheet	One-page summary table

---

2. Dataset Summary

Metric	Value
Total records	2,000
Total messages (all turns)	4,000 (2,000 × 2)
Turns per conversation	2 (1 human + 1 assistant)
Question script	Romanized Nepali (Latin letters)
Answer script	Devanagari (Nepali)
Unique companies referenced	210
Unique fiscal years referenced	18
Unique dividend types asked about	2 (total dividend, bonus share dividend)
Unique question phrasing templates	3 (combined with 2 dividend types → 6 total surface forms)
Unique dividend percentage values seen	392
Schema consistency (top-level)	100% — all 19 top-level keys present in every record
metadata_json schema consistency	100% — all 9 nested keys present in every record
generation_type / condition	real / real — sourced from actual dividend history, not synthetic text
License	Apache-2.0, permissive tier

---

3. File / Record Schema

Each line of the .jsonl file is a single, independent JSON object with the following 19 top-level fields:

Field	Type	Description	Example Value
id	string	Unique record identifier	"nepse_dividend_0001"
conversations	array of objects	The 1 question + 1 answer pair (see Section 5)	[{"from": "human", "value": "..."}, {"from": "gpt", "value": "..."}]
source	string	Human-readable name of the originating dataset	"NEPSE Dividend History"
source_name	string	Machine-friendly name of the source dataset	"nepse_dividend_faq"
source_repo	string	Repository/dataset name this record was pulled from	"NEPSE Dividend History"
source_config	string	Config name of the source dataset	"default"
source_split	string	Data split label	"train"
source_revision	string	Version tag of the source dataset	"v1"
source_row_id	string	Row identifier in the original source table (mirrors id)	"nepse_dividend_0001"
language	string	ISO 639-1 language code	"ne"
language_code	string	ISO 639-3 language code	"npi"
script	string	Unicode script code of the declared content	"Deva" (Devanagari)
license	string	License identifier for this data	"Apache-2.0"
license_tier	string	Human-readable license permissiveness classification	"permissive"
task_type	string	NLP task category this dataset supports	"instruction-following"
generation_type	string	How the record's content originated	"real" (factual/real data, not synthetically written)
condition	string	Content classification label	"real"
url	string	Source URL reference (empty for every record in this file)	""
metadata_json	string (JSON-encoded)	Nested metadata object — see Section 4	"{\"generation_domain\": \"वित्तीय सेवा\", ...}"

3.1 Nested schema: the conversations array

Each element of conversations is an object with exactly two fields:

Field	Type	Description	Allowed Values
from	string	Speaker role	"human" or "gpt"
value	string	The message text	Romanized Nepali (for human) or Devanagari Nepali (for gpt)

---

4. Nested metadata_json Schema

The metadata_json field is a JSON-encoded string (must be parsed with a second json.loads() call) containing 9 keys. These values are constant across all 2,000 records in this file — i.e., the entire dataset represents one single, narrowly-scoped generation configuration.

Nested Field	Type	Fixed Value (Nepali)	English Gloss
generation_domain	string	वित्तीय सेवा	Financial services
generation_category	string	नेप्से लाभांश इतिहास	NEPSE dividend history
generation_sub_domain	string	कम्पनी लाभांश	Company dividends
behavior	string	तथ्यात्मक लाभांश जानकारी दिने	Providing factual dividend information
behavior_definition	string	निर्दिष्ट कम्पनी र आर्थिक वर्षको लाभांश तथ्यांक प्रयोग गरी प्रश्नको प्रत्यक्ष उत्तर दिने।	"Give a direct answer to the question using the dividend figures for the specified company and fiscal year."
question_type	string	तथ्यमा आधारित प्रश्न	Fact-based question
content_language	string	नेपाली	Nepali
content_script	string	देवनागरी	Devanagari
english_content_allowed	boolean	false	English content is not permitted in the answer

Note: Although english_content_allowed is false and content_script declares Devanagari, the question side of every conversation is actually written in romanized Latin-script Nepali — the false/देवनागरी metadata describes the intended answer content policy, not the literal question text. This is an important nuance to document (see Section 16).

---

5. Conversation Structure

Unlike multi-turn conversational datasets, this dataset is strictly single-turn: every record contains exactly 2 messages.

Turn #	Index in array	Speaker (from)	Script	Purpose
1	conversations[0]	human	Romanized Nepali (Latin letters)	Asks for a specific company's dividend percentage for a specific fiscal year
2	conversations[1]	gpt	Devanagari Nepali	States the factual dividend percentage in a fixed sentence template

Structural Metric	Value
Messages per conversation	Always exactly 2
Role order	100% consistent: human, gpt
Human : assistant ratio	1:1

---

6. Question Pattern Diversity

Every question is built from one of 3 phrasing templates, applied to one of 2 dividend types, giving 6 total surface-form templates, each filled in with a specific company code and fiscal year. All 2,000 questions fall cleanly into these 6 buckets.

6.1 The 3 phrasing templates (using <CO> = company code, <FY> = fiscal year)

Template ID	Romanized Pattern	English Translation
A	aarthik barsha <FY> ma <CO> ko kul labhansha kati thiyo?	"In fiscal year <FY>, what was <CO>'s total dividend?"
B	<CO> le aarthik barsha <FY> ka lagi kul kati pratishat labhansha ghoshana garyo?	"What total dividend percentage did <CO> declare for fiscal year <FY>?"
C	<CO> ko aarthik barsha <FY> ko kul labhansha pratishat kati ho?	"What is <CO>'s total dividend percentage for fiscal year <FY>?"

For the bonus-share dividend type, the same 3 templates are used with kul labhansha ("total dividend") replaced by bonas seyar (labhansha) ("bonus share (dividend)"):

Template ID	Romanized Pattern	English Translation
A′	aarthik barsha <FY> ma <CO> ko bonas seyar labhansha kati thiyo?	"In fiscal year <FY>, what was <CO>'s bonus share dividend?"
B′	<CO> le aarthik barsha <FY> ka lagi kati pratishat bonas seyar ghoshana garyo?	"What bonus share percentage did <CO> declare for fiscal year <FY>?"
C′	<CO> ko aarthik barsha <FY> ko bonas seyar labhansha pratishat kati ho?	"What is <CO>'s bonus share dividend percentage for fiscal year <FY>?"

6.2 Distribution across the 6 question forms

Dividend Type	Template A (aarthik barsha <FY> ma ...)	Template B (<CO> le ... ghoshana garyo?)	Template C (<CO> ko ... pratishat kati ho?)	Subtotal
Total dividend (kul labhansha)	468	448	401	1,317
Bonus share dividend (bonas seyar)	231	221	231	683
Total	699	669	632	2,000

Metric	Value
Records asking about total dividend	1,317 (65.9%)
Records asking about bonus share dividend	683 (34.2%)
Most common single phrasing template	Template A, total-dividend form (468 records)
Least common single phrasing template	Template B, bonus-share form (221 records)

---

7. Answer Pattern Diversity

Unlike the free-form assistant answers in narrative health-style datasets, every answer here follows one single fixed sentence template, varying only in the company code, fiscal year, and numeric percentage:

Dividend Type	Answer Template
Total dividend	<CO> ko aarthik barsha <FY> ko kul labhansha <PCT>% thiyo. ("<CO>'s total dividend for fiscal year <FY> was <PCT>%.")
Bonus share dividend	<CO> ko aarthik barsha <FY> ko bonas seyar labhansha <PCT>% thiyo. ("<CO>'s bonus share dividend for fiscal year <FY> was <PCT>%.")

Metric	Value
Unique answer templates (structural forms)	2 (one per dividend type)
Unique answer strings (with company/year/percentage filled in)	2,000 (each is effectively unique due to varying numeric percentage)
Answer always states the company code, fiscal year, and percentage	100% of records
Answer includes any disclaimer or hedge language	None — every answer is a direct factual statement (no safety caveat, consistent with question_type: तथ्यमा आधारित प्रश्न / "fact-based question")

---

8. Company Coverage

The dataset references 210 unique NEPSE-listed companies, each identified by a romanized letter-by-letter spelling of its stock ticker symbol (e.g. the ticker "SLBSL" is spelled out phonetically as esielabiesaela, "S-L-B-S-L"). Company names are not given in expanded/full form anywhere in the dataset — only these phonetic ticker spellings are used.

Metric	Value
Unique companies (romanized ticker codes)	210
Minimum records for a single company	1
Maximum records for a single company	32
Average records per company	~9.5

8.1 Most frequently referenced companies (top 15 by record count)

#	Romanized Ticker Code	Record Count
1	ibiela	32
2	jibibiela	31
3	sibibiela	30
4	siechasiela	30
5	siaaiti	30
6	didibiela	30
7	jibiaaiemai	30
8	aaisiephasi	30
9	emadibi	30
10	enaebiaaiela	30
11	sijedabiaaiela	29
12	bipisiela	28
13	ephaemadibiela	28
14	eechapisi	26
15	echadiela	26

The remaining ~195 companies appear with lower frequency, down to companies referenced only once. Because tickers are spelled phonetically letter-by-letter rather than given as standard stock symbols, mapping a given code back to a real-world company name requires manually decoding each letter (e.g., e = "E", bi = "B", si = "C", di = "D", ela = "L", etc.) — see Section 16 for a caveat on this.

---

9. Fiscal Year Coverage

Fiscal years are expressed in the Bikram Sambat (BS) calendar, in YYYY/YYYY format (e.g. 2077/2078 corresponds roughly to 2020/2021 in the Gregorian calendar). The dataset spans 18 unique fiscal years.

Fiscal Year (BS)	Record Count
2062/2063	2
2066/2067	1
2067/2068	67
2068/2069	75
2069/2070	74
2070/2071	92
2071/2072	99
2072/2073	116
2073/2074	129
2074/2075	131
2075/2076	149
2076/2077	160
2077/2078	178
2078/2079	180
2079/2080	134
2080/2081	172
2081/2082	206
2082/2083	35

Metric	Value
Earliest fiscal year	2062/2063
Latest fiscal year	2082/2083
Most represented fiscal year	2081/2082 (206 records — the most recent full fiscal year, likely reflecting more complete recent reporting)
Least represented fiscal year	2066/2067 (1 record)
General trend	Coverage increases steadily from the oldest years toward 2078/2079, dips slightly in 2079/2080, then rises again — reflecting more companies having reported dividends in more recent years, with 2082/2083 (still in progress) being comparatively sparse

---

10. Dividend Value Coverage

Dividend percentages are stated as plain numbers followed by %, and include both whole numbers and precise decimal values (a result of real-world dividend calculations, e.g. splitting a cash dividend across share counts).

Metric	Value
Unique dividend percentage values across dataset	392
Most common dividend value	10% (124 records)
2nd most common	15% (120 records)
3rd most common	0% (114 records — companies that declared no dividend that year)
4th most common	20% (102 records)
5th most common	5% (85 records)

10.1 Top 15 most frequent dividend values

Dividend Value	Count
10%	124
15%	120
0%	114
20%	102
5%	85
10.53%	68
25%	55
21.05%	52
15.79%	48
5.26%	46
8%	37
12%	37
30%	36
7%	30
14%	29

The presence of 0% as the third most common value (114 records) is notable — it correctly represents years in which a company declared no dividend, confirming this is real, unfiltered historical data rather than a cherry-picked "success stories only" dataset.

---

11. Length & Size Statistics

All lengths measured in characters.

Measure	Minimum	Maximum	Mean
Human (question) message length	60	101	79.7
Assistant (gpt) answer message length	58	82	70.2

Structural Metric	Value
Messages per conversation	Always exactly 2
Total messages in dataset	4,000
Records with missing/malformed fields	0
Records with non-2 turn count	0

---

12. Metadata Field Value Reference

Quick reference for every metadata field that is constant across the dataset (useful for filtering scripts, even though the value never varies here):

Field	Fixed Value
source	NEPSE Dividend History
source_name	nepse_dividend_faq
source_repo	NEPSE Dividend History
source_config	default
source_split	train
source_revision	v1
language	ne
language_code	npi
script	Deva
license	Apache-2.0
license_tier	permissive
task_type	instruction-following
generation_type	real
condition	real
url	"" (always empty)
metadata_json.generation_domain	वित्तीय सेवा
metadata_json.generation_category	नेप्से लाभांश इतिहास
metadata_json.generation_sub_domain	कम्पनी लाभांश
metadata_json.behavior	तथ्यात्मक लाभांश जानकारी दिने
metadata_json.question_type	तथ्यमा आधारित प्रश्न
metadata_json.content_language	नेपाली
metadata_json.content_script	देवनागरी
metadata_json.english_content_allowed	false

Fields that vary per record:

Field	Range / Cardinality
id	nepse_dividend_0001 → nepse_dividend_2000
source_row_id	Identical to id for every record
conversations	2,000 distinct question/answer pairs
Company referenced (embedded in text)	210 distinct romanized ticker codes
Fiscal year referenced (embedded in text)	18 distinct BS fiscal years
Dividend percentage (embedded in text)	392 distinct values
Dividend type (embedded in text)	2 (total dividend / bonus share dividend)

---

13. Loading the Dataset (Code Demos)

13.1 Plain Python — streaming line-by-line (no dependencies)

import json

path = "nepse_dividend_faq_romanized.jsonl"

records = []
with open(path, encoding="utf-8") as f:
    for line in f:
        line = line.strip()
        if not line:
            continue
        record = json.loads(line)
        records.append(record)

print(f"Loaded {len(records)} records")

# Print the first Q&A pair, plus its parsed nested metadata
first = records[0]
question = first["conversations"][0]["value"]
answer = first["conversations"][1]["value"]
print("Q:", question)
print("A:", answer)

metadata = json.loads(first["metadata_json"])  # metadata_json is a JSON-encoded string
print("Behavior:", metadata["behavior"])

13.2 Using pandas

import pandas as pd
import json

path = "nepse_dividend_faq_romanized.jsonl"

df = pd.read_json(path, lines=True)
print(df.shape)              # (2000, 19)
print(df.columns.tolist())   # all 19 top-level fields

# Split question/answer into their own columns
df["question"] = df["conversations"].apply(lambda c: c[0]["value"])
df["answer"] = df["conversations"].apply(lambda c: c[1]["value"])

# Parse the nested metadata_json string into real columns
meta_df = df["metadata_json"].apply(json.loads).apply(pd.Series)
df = pd.concat([df.drop(columns=["conversations", "metadata_json"]), df[["question", "answer"]], meta_df], axis=1)

df.to_csv("nepse_dividend_flat.csv", index=False, encoding="utf-8-sig")
print(df.head())

13.3 Extracting structured (company, fiscal_year, dividend_type, percentage) tuples

import json
import re

fy_pattern = re.compile(r"20\d{2}/20\d{2}")
pct_pattern = re.compile(r"\d+(?:\.\d+)?%")

def parse_answer(answer_text: str):
    """Extract company code, fiscal year, dividend type, and percentage from a gpt answer."""
    fiscal_year = fy_pattern.search(answer_text).group()
    percentage = pct_pattern.search(answer_text).group()
    dividend_type = "bonus_share" if "bonas seyar" in answer_text else "total"
    company_code = answer_text.split(" ko aarthik barsha")[0]
    return {
        "company_code": company_code,
        "fiscal_year": fiscal_year,
        "dividend_type": dividend_type,
        "percentage": percentage,
    }

with open("nepse_dividend_faq_romanized.jsonl", encoding="utf-8") as f:
    records = [json.loads(line) for line in f]

structured = [parse_answer(r["conversations"][1]["value"]) for r in records]
print(structured[:5])

13.4 Using Hugging Face datasets

from datasets import load_dataset

dataset = load_dataset(
    "json",
    data_files="nepse_dividend_faq_romanized.jsonl",
    split="train",
)

print(dataset)
print(dataset[0]["conversations"])

# Filter to records about bonus share dividends only
bonus_share = dataset.filter(
    lambda r: "bonas seyar" in r["conversations"][0]["value"]
)
print(len(bonus_share))  # 683

# Filter to a single fiscal year
fy_2081_82 = dataset.filter(
    lambda r: "2081/2082" in r["conversations"][0]["value"]
)
print(len(fy_2081_82))  # 206

13.5 Converting to a chat-fine-tuning message format

import json

def to_chat_messages(record):
    role_map = {"human": "user", "gpt": "assistant"}
    return [
        {"role": role_map[turn["from"]], "content": turn["value"]}
        for turn in record["conversations"]
    ]

with open("nepse_dividend_faq_romanized.jsonl", encoding="utf-8") as f:
    records = [json.loads(line) for line in f]

chat_formatted = [to_chat_messages(r) for r in records]

with open("nepse_dividend_chat_format.jsonl", "w", encoding="utf-8") as out:
    for messages in chat_formatted:
        out.write(json.dumps({"messages": messages}, ensure_ascii=False) + "\n")

---

14. Example Records

14.1 Total dividend example

{
  "id": "nepse_dividend_0001",
  "conversations": [
    {"from": "human", "value": "aarthik barsha 2077/2078 ma esielabiesaela ko kul labhansha kati thiyo?"},
    {"from": "gpt", "value": "esielabiesaela ko aarthik barsha 2077/2078 ko kul labhansha 21% thiyo."}
  ],
  "source": "NEPSE Dividend History",
  "source_name": "nepse_dividend_faq",
  "source_repo": "NEPSE Dividend History",
  "source_config": "default",
  "source_split": "train",
  "source_revision": "v1",
  "source_row_id": "nepse_dividend_0001",
  "language": "ne",
  "language_code": "npi",
  "script": "Deva",
  "license": "Apache-2.0",
  "license_tier": "permissive",
  "task_type": "instruction-following",
  "generation_type": "real",
  "condition": "real",
  "url": "",
  "metadata_json": "{\"generation_domain\": \"वित्तीय सेवा\", \"generation_category\": \"नेप्से लाभांश इतिहास\", \"generation_sub_domain\": \"कम्पनी लाभांश\", \"behavior\": \"तथ्यात्मक लाभांश जानकारी दिने\", \"behavior_definition\": \"निर्दिष्ट कम्पनी र आर्थिक वर्षको लाभांश तथ्यांक प्रयोग गरी प्रश्नको प्रत्यक्ष उत्तर दिने।\", \"question_type\": \"तथ्यमा आधारित प्रश्न\", \"content_language\": \"नेपाली\", \"content_script\": \"देवनागरी\", \"english_content_allowed\": false}"
}

14.2 Bonus share dividend example

{
  "id": "nepse_dividend_0027",
  "conversations": [
    {"from": "human", "value": "esielabiesaela ko aarthik barsha 2077/2078 ko bonas seyar labhansha pratishat kati ho?"},
    {"from": "gpt", "value": "esielabiesaela ko aarthik barsha 2077/2078 ko bonas seyar labhansha 21% thiyo."}
  ]
}

(Non-varying metadata fields omitted for brevity — see Section 12 for the full fixed values.)

---

15. Suggested Use Cases

Use Case	Why This Dataset Fits
Fine-tuning a Nepali financial-FAQ assistant	Clean, real, factual Q&A pairs about a well-defined financial topic (dividends)
Romanized-input → Devanagari-output transliteration-aware QA	Questions are in Latin-script romanized Nepali; answers are in Devanagari — directly models this input/output script mismatch
Structured information extraction research	Every question/answer can be reduced to a clean 4-tuple: (company, fiscal_year, dividend_type, percentage)
Evaluating factual consistency / hallucination in financial QA	Ground-truth answers are deterministic and verifiable against real dividend history
Building a NEPSE dividend lookup chatbot	210 companies × 18 fiscal years × 2 dividend types gives broad, realistic query coverage
Named-entity recognition for stock ticker codes (romanized)	210 unique phonetically-spelled ticker codes provide a distinctive NER challenge
Studying real-world data sparsity/reporting patterns	The declining/uneven per-year record counts (Section 9) reflect genuine reporting availability, not synthetic balancing

---

16. Limitations & Notes

Point	Detail
Real, not synthetic	generation_type and condition are both real — unlike template-generated conversational datasets, the underlying dividend facts come from actual NEPSE company history. However, the question phrasing is still templated (6 fixed surface forms), so linguistic diversity is low even though factual content is real.
Romanized company codes are letter-by-letter ticker spellings, not real names	Codes like esielabiesaela or edibiela are phonetic spellings of stock ticker symbols (e.g., "S-L-B-S-L", "D-B-L"). The dataset does not include a mapping table from these codes to full company names — users wanting real company names will need to manually decode the phonetic spelling or cross-reference an external NEPSE ticker list.
No date of dividend declaration, only fiscal year	The dataset gives the Bikram Sambat fiscal year (e.g. 2077/2078) but not the exact announcement date, AGM date, or book-closure date.
0% entries represent genuine "no dividend" years, not missing data	114 records with 0% reflect real years where a company declared no dividend — this is a meaningful signal, not a placeholder.
Only 2 dividend categories covered	Only "total dividend" and "bonus share dividend" are represented; the dataset does not separately break out cash dividend vs. stock dividend components in all cases, nor does it cover rights shares, book value, EPS, or other financial metrics.
Uneven company/year coverage	Some companies appear up to 32 times, others only once; some fiscal years have over 200 records, others as few as 1–2 — this reflects real data availability, not an intentionally balanced sample.
No demographic/personal identifiers	This is company-level financial history; no personal or user data is present.
url field is empty for all records	No direct source URL is provided per record; provenance is only available at the dataset level (source, source_repo, source_revision).

---

17. Quick Reference Cheat Sheet

Item	Value
File name	nepse_dividend_faq_romanized.jsonl
Format	JSON Lines
Total records	2,000
Messages per record	2 (1 human + 1 assistant)
Total messages	4,000
Question script	Romanized Nepali (Latin)
Answer script	Devanagari (Nepali)
Domain	Financial services — NEPSE dividend history
Data nature	Real (factual), not synthetic
Fields per record (top-level)	19
Nested metadata fields	9
Schema consistency	100%
Unique companies	210
Unique fiscal years	18 (2062/2063 → 2082/2083)
Unique dividend types	2 (total / bonus share)
Unique question phrasing templates	3 (× 2 dividend types = 6 surface forms)
Unique dividend percentage values	392
Avg. question length	79.7 characters
Avg. answer length	70.2 characters
License	Apache-2.0 (permissive)