sabin1234/Nepali_Finance_FAQ_Dataset

Nepali Finance FAQ Dataset (merged_finance_faq_serial.jsonl)

A merged, serially-indexed collection of real, institution-sourced Frequently Asked Questions (FAQ) pairs from Nepal's banking, insurance, mutual fund, and personal-finance domain — written entirely in Nepali (Devanagari script) — packaged in instruction-tuning (human / gpt) conversational format.

---

1. Overview

Property	Value
File name	merged_finance_faq_serial.jsonl
Format	JSON Lines (.jsonl) — one JSON object per line
Total records	391
File size	~816 KB
Language	Nepali (ne / ISO 639-3 npi)
Script	Devanagari (Deva)
License	Apache-2.0 (permissive)
Task type	Instruction-following (single-turn Q&A)
Data origin	Real, publicly published FAQ content ("real" generation type — not synthetically generated)
Number of source institutions	13
Conversation structure	Every record has exactly 2 turns: human (question) → gpt (answer)

This dataset was built by merging FAQ pages from 13 different Nepali financial institutions and public bodies into one unified schema, then assigning each row a single continuous serial ID (000001 → 000391) while preserving each source's own local ID prefix.

---

2. Why this dataset exists

Nepali is a low-resource language for finance-domain NLP. Public FAQ pages from banks, insurers, and financial regulators in Nepal are a natural, high-quality source of factual, domain-grounded, real-world question–answer pairs — ideal for:

• Fine-tuning or instruction-tuning LLMs to answer Nepali financial questions accurately.
• Building retrieval-augmented generation (RAG) knowledge bases for Nepali fintech/insurtech chatbots.
• Evaluating LLMs on Nepali reading comprehension / factual QA in a specialized domain.
• Research on low-resource-language financial NLP.

---

3. Row / Record Schema

Every line in the .jsonl file is a single JSON object with the following top-level fields:

Field	Type	Description
id	string	Globally unique record ID, e.g. siddhartha_faq_000241. Source-specific prefix + a 6-digit serial number that is continuous across the whole file (000001–000391).
conversations	array[object]	The Q&A turn pair (see §4 below). Always length 2: [human, gpt].
source	string	Human-readable name of the originating institution, in Nepali (e.g. सिद्धार्थ बैंक).
source_name	string	Machine-friendly snake_case identifier of the source dataset, e.g. siddhartha_bank_faq.
source_repo	string	Name of the originating organization in English, e.g. Siddhartha Bank.
source_config	string	Dataset configuration name (always default in this file).
source_split	string	Data split the row belongs to (always train).
source_revision	string	Version tag of the source data (always v1).
source_row_id	string	The row's ID within its original (pre-merge) source file — identical to id in this merged file.
language	string	ISO 639-1 language code — always ne (Nepali).
language_code	string	ISO 639-3 language code — always npi.
script	string	Unicode script name — always Deva (Devanagari).
license	string	License of the underlying content — always Apache-2.0.
license_tier	string	License category — always permissive.
task_type	string	NLP task category — always instruction-following.
generation_type	string	How the content was produced — always real (i.e., sourced from real published FAQ text, not LLM-synthesized).
condition	string	Data-collection condition — always real.
url	string	Source URL for the FAQ entry (empty string "" for every row in this file — no URLs were retained during merging).
metadata_json	string	A JSON-encoded string (not a nested object) holding fine-grained annotation metadata — see §5 below. You must json.loads() this field separately to access its contents.

4. conversations field structure

"conversations": [
  { "from": "human", "value": "<the question, in Nepali>" },
  { "from": "gpt",   "value": "<the answer, in Nepali>" }
]

• from is always either "human" (the FAQ question) or "gpt" (the FAQ answer).
• Every single record in the dataset follows this exact 2-turn pattern — there are no multi-turn dialogues, no system turns, and no missing turns.

5. metadata_json field structure (after parsing)

metadata_json is a string containing an embedded JSON object. Once parsed with json.loads(), it exposes:

Key	Type	Description
generation_domain	string	Top-level subject domain — always वित्त ("Finance").
generation_category	string	The specific FAQ set / institution category the row came from, in Nepali (e.g. सिद्धार्थ बैंक सम्बन्धी प्रायः सोधिने प्रश्न — "Frequently asked questions about Siddhartha Bank").
generation_sub_domain	string	Finer-grained topical sub-domain (e.g. बैंकिङ सेवा — "Banking services", म्युचुअल फन्ड लगानी — "Mutual fund investment").
behavior	string	The intended model behavior label — always जानकारीमूलक उत्तर दिने ("Provide an informative answer").
behavior_definition	string	A one-sentence Nepali description of that behavior: preserve the meaning of the original question and the facts of the original answer while giving the response in Nepali.
question_type	string	Question category — always तथ्यमा आधारित प्रश्न ("Fact-based question").
content_language	string	Content language, Nepali-language label — always नेपाली.
content_script	string	Content script, Nepali-language label — always देवनागरी.
english_content_allowed	boolean	Whether English content is permitted in the answer — always false for this file (answers are Nepali-only).

Important diversity note: Within this particular merged file, the annotation-level fields behavior, behavior_definition, question_type, and generation_domain are uniform across all 391 rows (single value each — see the tables in §7). This is expected: the whole file represents one behavior class ("give a faithful, informative Nepali answer to a fact-based finance question"). The actual diversity in this dataset lives at three other levels instead:

1. Source institution diversity — 13 distinct organizations (§6).
2. Sub-domain / topical diversity — 10 distinct finance sub-domains (§7.3).
3. Surface-level question-pattern diversity — the literal phrasing/structure of the Nepali questions themselves varies a lot (what/how/how-much/where/why/can-I style questions, etc. — see §8).

---

6. Source Institutions

The dataset merges FAQ content from the following 13 Nepali financial entities. Row counts and ID ranges are contiguous per source (the file is grouped by source, not shuffled):

#	source (Nepali)	source_repo (English)	source_name	Row count	ID prefix example	Serial ID range
1	नेपाल बीमा प्राधिकरण	Nepal Insurance Authority	nia_faq	21	nia_######	nia_000001 – nia_000021
2	—	Nepal Investment Mega Bank (NIMB Smart)	nimb_smart_faq	9	nimb_smart_faq_######	..._000022 – ..._000030
3	—	National Institute of Securities Markets (Mutual Funds)	nism_mutual_fund_faq	50	nism_mf_######	nism_mf_000031 – nism_mf_000080
4	—	NLG Insurance Company Ltd.	nlg_insurance_faq	19	nlg_######	nlg_000081 – nlg_000099
5	—	NMB Bank	nmb_bank_faq	20	nmb_faq_######	nmb_faq_000100 – nmb_faq_000119
6	—	Prime Commercial Bank	prime_commercial_bank_faq	59	prime_bank_faq_######	..._000120 – ..._000178
7	—	Reliance Finance Limited	reliancefinance_faq	11	reliancefinance_######	..._000179 – ..._000189
8	—	Sanima GIC Insurance Ltd.	sanima_gic_faq	7	sanima_gic_######	..._000190 – ..._000196
9	—	Shine Resunga Development Bank	shine_resunga_faq	44	srdb_######	srdb_000197 – srdb_000240
10	—	Siddhartha Bank	siddhartha_bank_faq	81	siddhartha_faq_######	..._000241 – ..._000321
11	—	Tata Moneyfy	tata_moneyfy_faq	30	tata_moneyfy_faq_######	..._000322 – ..._000351
12	—	WealthSerene (personal finance blog)	wealthserene_faq	20	wealthserene_######	..._000352 – ..._000371
13	—	World Bank (external debt statistics)	world_bank_faq	20	worldbank_faq_######	..._000372 – ..._000391

Total: 391 rows across 13 sources.

Institution types represented:

• Commercial / development banks: Siddhartha Bank, Prime Commercial Bank, NMB Bank, Shine Resunga Development Bank, Nepal Investment Mega Bank (NIMB) — mobile/internet banking, account opening, cards, loans.
• Regulators / public institutions: Nepal Insurance Authority (नेपाल बीमा प्राधिकरण), National Institute of Securities Markets, World Bank.
• Insurance companies: NLG Insurance, Sanima GIC Insurance.
• NBFC / finance companies: Reliance Finance Limited.
• Fintech / personal finance: Tata Moneyfy (loan marketplace), WealthSerene (personal-finance education blog).

---

7. Distributional Statistics

7.1 Fields that are constant across all 391 rows

These fields take exactly one value for every single record in the file (no variation):

Field	Constant value
language	ne
language_code	npi
script	Deva
license	Apache-2.0
license_tier	permissive
task_type	instruction-following
generation_type	real
condition	real
source_config	default
source_split	train
source_revision	v1
url	"" (empty)
metadata_json.generation_domain	वित्त ("Finance")
metadata_json.behavior	जानकारीमूलक उत्तर दिने ("Provide an informative answer")
metadata_json.question_type	तथ्यमा आधारित प्रश्न ("Fact-based question")
metadata_json.content_language	नेपाली
metadata_json.content_script	देवनागरी
metadata_json.english_content_allowed	false
Number of conversations turns	2 (always human then gpt)

7.2 Row counts per source (behavior/question-set distribution)

Source (source_name)	Rows	% of dataset
siddhartha_bank_faq	81	20.7%
prime_commercial_bank_faq	59	15.1%
nism_mutual_fund_faq	50	12.8%
shine_resunga_faq	44	11.3%
tata_moneyfy_faq	30	7.7%
nia_faq	21	5.4%
nmb_bank_faq	20	5.1%
wealthserene_faq	20	5.1%
world_bank_faq	20	5.1%
nlg_insurance_faq	19	4.9%
reliancefinance_faq	11	2.8%
nimb_smart_faq	9	2.3%
sanima_gic_faq	7	1.8%
Total	391	100%

7.3 Sub-domain (generation_sub_domain) distribution

This is where most of the topical diversity in the dataset lives:

Sub-domain (Nepali)	English gloss	Rows
बैंकिङ सेवा	Banking services	204
म्युचुअल फन्ड लगानी	Mutual fund investment	50
ऋण र म्युचुअल फन्ड	Loans & mutual funds	30
निर्जीवन बीमा	Non-life (general) insurance	26
बीमा नियमन	Insurance regulation	21
बजेटिङ र आपतकालीन कोष	Budgeting & emergency funds	20
वित्तीय सेवा	Financial services	11
ऋण तथ्याङ्क र पारदर्शिता	Debt statistics & transparency	10
ऋण तथ्याङ्क र धोखाधडी जागरूकता	Debt statistics & fraud awareness	10
मोबाइल बैंकिङ	Mobile banking	9
Total		391

7.4 Per-source FAQ category (generation_category) list

Each source maps to exactly one category label (13 distinct categories, matching the 13 sources one-to-one):

Category (Nepali)	English gloss	Rows
सिद्धार्थ बैंक सम्बन्धी प्रायः सोधिने प्रश्न	FAQs about Siddhartha Bank	81
प्राइम कमर्सियल बैंक सम्बन्धी प्रायः सोधिने प्रश्न	FAQs about Prime Commercial Bank	59
म्युचुअल फन्ड सम्बन्धी प्रायः सोधिने प्रश्न	FAQs about mutual funds	50
शाइन रेसुङ्गा डेभलपमेन्ट बैंक सम्बन्धी प्रायः सोधिने प्रश्न	FAQs about Shine Resunga Development Bank	44
टाटा मनिफाई सम्बन्धी प्रायः सोधिने प्रश्न	FAQs about Tata Moneyfy	30
नेपाल बीमा प्राधिकरण सम्बन्धी प्रायः सोधिने प्रश्न	FAQs about Nepal Insurance Authority	21
एनएमबी बैंक सम्बन्धी प्रायः सोधिने प्रश्न	FAQs about NMB Bank	20
वेल्थसेरीन व्यक्तिगत वित्त सम्बन्धी प्रायः सोधिने प्रश्न	FAQs about WealthSerene personal finance	20
विश्व बैंक सम्बन्धी प्रायः सोधिने प्रश्न	FAQs about the World Bank	20
एनएलजी इन्स्योरेन्स सम्बन्धी प्रायः सोधिने प्रश्न	FAQs about NLG Insurance	19
रिलायन्स फाइनान्स सम्बन्धी प्रायः सोधिने प्रश्न	FAQs about Reliance Finance	11
नेपाल इन्भेस्टमेन्ट मेगा बैंक स्मार्ट सम्बन्धी प्रायः सोधिने प्रश्न	FAQs about NIMB Smart (Nepal Investment Mega Bank)	9
सनिमा जीआईसी इन्स्योरेन्स सम्बन्धी प्रायः सोधिने प्रश्न	FAQs about Sanima GIC Insurance	7

7.5 Text length statistics

Metric	Min	Max	Average
Question length (characters)	11	134	~49.3
Answer length (characters)	36	699	~221.2

Answers are, on average, ~4.5x longer than questions — consistent with genuine FAQ content where questions are short prompts and answers are explanatory paragraphs.

---

8. Question Pattern Diversity

Although the annotation-level question_type field is uniform (तथ्यमा आधारित प्रश्न, "fact-based question"), the surface phrasing of the 391 questions is genuinely varied. The table below groups the most common Nepali question-opening patterns found in the human turns, with their English meaning and a real example from the dataset:

Question pattern (Nepali opener)	English meaning of pattern	Occurrences (approx., top openers)	Example question (Nepali)	English gloss
के ... ?	"What is / Is it ...?" — definitional & yes/no questions	48	नेपाल बीमा प्राधिकरण भनेको के हो?	"What is the Nepal Insurance Authority?"
मेरो ...	"My ..." — personal/account-context questions	28	मेरो खाता कसरी हेर्ने? (style)	"How do I view my account?"
के म ... सक्छु?	"Can I ...?" — eligibility / permission questions	(subset of के)	के म एनएमबी मोबाइल बैंकिङ सेवा सदस्य बन्न सक्छु?	"Can I become a member of NMB mobile banking service?"
कसरी ... ?	"How do I / How to ...?" — procedural questions	frequent across sources	अनलाइन खाता कसरी खोल्ने?	"How to open an online account?"
कति ... ?	"How much / how many ...?" — quantity questions	frequent	बजारमा कति प्रकारका बीमा छन्?	"How many types of insurance are there in the market?"
कहाँ ... ?	"Where ...?" — location questions	occasional	नेपाल बीमा प्राधिकरणको कार्यालय कहाँ रहेको छ?	"Where is the office of the Nepal Insurance Authority located?"
किन ... ?	"Why ...?" — rationale questions	occasional	सिद्धार्थ बैंकस्मार्ट एक्सपी के हो? किन आवेदन दिने?	"What is Siddhartha BankSmart XP? Why apply?"
[Product/brand name] + के हो?	"What is [Product]?" — product-definition questions, brand-name led	very common (per-source variant of "के")	निम्ब स्मार्ट के हो?	"What is NIMB Smart?"
[Institution name]-ले के काम गर्दछ?	"What does [Institution] do?" — functional/role questions	occasional	नेपाल बीमा प्राधिकरणले के काम गर्दछ?	"What does the Nepal Insurance Authority do?"
को हो?	"Who is ...?" — role/identity questions	occasional	लाभार्थी मालिक (बीओ) को हो?	"Who is the Beneficial Owner (BO)?"

Broad behavioral shape of the answers: every gpt turn is a single, self-contained explanatory paragraph — no follow-up questions, no bullet-formatted answers, no refusals, no multi-turn back-and-forth. This is uniform across the dataset and reflects the source data being static, published FAQ pages rather than live chat transcripts.

---

9. Representative Examples (one per source, verbatim from the file)

Nepali text is shown as-is; an English gloss (not part of the dataset) is added by this README for readability only.

Source	ID	Question (Nepali)	Gloss	Answer (Nepali, truncated)	Gloss
Nepal Insurance Authority	nia_000001	नेपाल बीमा प्राधिकरण भनेको के हो?	What is the Nepal Insurance Authority?	बीमा ऐन, २०७९ बमोजिम ... नेपालको बीमा नियमनकारी निकाय हो।	It is Nepal's insurance regulatory body, established under the Insurance Act 2079...
NIMB Smart (Nepal Investment Mega Bank)	nimb_smart_faq_000022	निम्ब स्मार्ट के हो?	What is NIMB Smart?	निम्ब स्मार्ट अनलाइन उपलब्ध सुविधाजनक र सुरक्षित कारोबार बैंकिङ एप हो।	NIMB Smart is a convenient and secure online transaction banking app...
NISM (Mutual Funds)	nism_mf_000032	म्युचुअल फन्ड के हो?	What is a mutual fund?	म्युचुअल फन्ड व्यावसायिक फन्ड व्यवस्थापकद्वारा व्यवस्थापन गरिने रकमको पोखरी हो।	A mutual fund is a pool of money managed by a professional fund manager...
NLG Insurance	nlg_000081	बीमा के हो?	What is insurance?	बीमा दुई पक्षबीचको करार हो जसमा ... क्षतिपूर्ति दिने आश्वासन दिन्छ।	Insurance is a contract between two parties in which one indemnifies the other...
NMB Bank	nmb_faq_000100	एनएमबी मोबाइल बैंकिङ सेवा के हो?	What is NMB Mobile Banking?	एनएमबी मोबाइल बैंकिङ एनएमबी बैंकले आफ्ना ग्राहकलाई प्रदान गर्ने नवीनतम मूल्य-वर्धित सेवा हो।	NMB Mobile Banking is a value-added service offered by NMB Bank to its customers...
Prime Commercial Bank	prime_bank_faq_000120	डिपोजिटरी सहभागी (डीपी) के/को हो?	What is a Depository Participant (DP)?	डिपोजिटरी सहभागी सेबोन र सीडीएससीद्वारा इजाजतपत्र प्राप्त संस्था हो ...	A DP is an entity licensed by SEBON and CDSC...
Reliance Finance Ltd.	reliancefinance_000179	रिलायन्स फाइनान्स लिमिटेडमा खाता खोल्न न्यूनतम ब्यालेन्स कति चाहिन्छ?	What is the minimum balance to open an account at Reliance Finance?	रिलायन्स फाइनान्स लिमिटेडमा खाता खोल्न न्यूनतम ब्यालेन्स रु. १०० चाहिन्छ।	The minimum balance required is Rs. 100.
Sanima GIC Insurance	sanima_gic_000190	बीमा के हो?	What is insurance?	बीमा दुई पक्षबीचको करार हो — बीमक र बीमित ...	Insurance is a contract between two parties — the insurer and the insured...
Shine Resunga Dev. Bank	srdb_000198	अनलाइन खाता कसरी खोल्ने?	How to open an online account?	अनलाइन खाता खोल्न बैंकको आधिकारिक वेबसाइट वा अनलाइन खाता खोल्ने पोर्टलबाट अगाडि बढ्न सकिन्छ।	You can proceed via the bank's official website or online account-opening portal...
Siddhartha Bank	siddhartha_faq_000241	सिद्धार्थ बैंकस्मार्ट एक्सपी के हो? किन आवेदन दिने?	What is Siddhartha BankSmart XP? Why apply?	सिद्धार्थ बैंकस्मार्ट एक्सपी सिद्धार्थ बैंकले प्रदान गर्ने सुविधाजनक र सुरक्षित मोबाइल बैंकिङ सेवा हो ...	Siddhartha BankSmart XP is a convenient, secure mobile banking service from Siddhartha Bank...
Tata Moneyfy	tata_moneyfy_faq_000322	मनिफाईको ऋणका लागि व्यवसायिक मोडेल के हो?	What is Moneyfy's business model for loans?	मनिफाईले आफ्ना ग्राहकलाई विभिन्न ऋण प्रदान गर्न बैंक र एनबीएफसीलाई आफ्नो प्लेटफर्ममा ल्याउँछ।	Moneyfy brings banks and NBFCs onto its platform to offer customers various loans...
WealthSerene	wealthserene_000352	आपतकालीन कोषमा कति रकम हुनुपर्छ?	How much should be in an emergency fund?	धेरैजसो वित्तीय शिक्षाले आवश्यक जीवनयापन खर्चको ३–६ महिना सिफारिस गर्दछ ...	Most financial education recommends 3–6 months of essential living expenses...
World Bank	worldbank_faq_000372	यो ऋण वेबसाइटले के प्रदान गर्दछ?	What does this debt website provide?	यो वेबसाइट ऋण तथ्याङ्क तथा सम्बन्धित प्रकाशन र समाचारको केन्द्रीय हब हो।	This website is a central hub for debt statistics and related publications...

---

10. Full JSON Record Example (raw, unmodified)

{
  "id": "nia_000001",
  "conversations": [
    {
      "from": "human",
      "value": "नेपाल बीमा प्राधिकरण भनेको के हो?"
    },
    {
      "from": "gpt",
      "value": "बीमा ऐन, २०७९ बमोजिम बीमा प्रणाली तथा बीमा व्यवसायलाई व्यवस्थित, नियमित, प्रतिस्पर्धी तथा विश्वसनीय बनाई त्यसको विकास गर्न, बीमा व्यवसायको स्वस्थ प्रतिस्पर्धाबाट सर्वसाधारणलाई गुणस्तरीय तथा भरपर्दो बीमा सेवा उपलब्ध गराउन एवम् बीमा व्यवसायको प्रभावकारी रूपमा नियमन गरी बीमितको हकहित संरक्षण गर्नको लागि स्थापना भएको नेपालको बीमा नियमनकारी निकाय हो।"
    }
  ],
  "source": "नेपाल बीमा प्राधिकरण",
  "source_name": "nia_faq",
  "source_repo": "Nepal Insurance Authority",
  "source_config": "default",
  "source_split": "train",
  "source_revision": "v1",
  "source_row_id": "nia_000001",
  "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\": \"वित्त\", \"generation_category\": \"नेपाल बीमा प्राधिकरण सम्बन्धी प्रायः सोधिने प्रश्न\", \"generation_sub_domain\": \"बीमा नियमन\", \"behavior\": \"जानकारीमूलक उत्तर दिने\", \"behavior_definition\": \"मूल प्रश्नको अर्थ र मूल उत्तरको तथ्य कायम राख्दै नेपालीमा रूपान्तरण गरिएको उत्तर दिने।\", \"question_type\": \"तथ्यमा आधारित प्रश्न\", \"content_language\": \"नेपाली\", \"content_script\": \"देवनागरी\", \"english_content_allowed\": false}"
}

Note that metadata_json is a string, not a nested JSON object — you must call json.loads() on it a second time after parsing the outer line.

---

11. How to Load the Dataset

11.1 Plain Python (standard library only)

import json

records = []
with open("merged_finance_faq_serial.jsonl", encoding="utf-8") as f:
    for line in f:
        line = line.strip()
        if not line:
            continue
        obj = json.loads(line)
        # metadata_json is itself a JSON string -> parse it too
        obj["metadata"] = json.loads(obj["metadata_json"])
        records.append(obj)

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

# Example: print the first question/answer pair
first = records[0]
question = first["conversations"][0]["value"]
answer = first["conversations"][1]["value"]
print("Source :", first["source_repo"])
print("Q:", question)
print("A:", answer)
print("Sub-domain:", first["metadata"]["generation_sub_domain"])

11.2 Using pandas

import json
import pandas as pd

# pandas can read JSONL directly with lines=True
df = pd.read_json("merged_finance_faq_serial.jsonl", lines=True)

# Expand the nested metadata_json string column into real columns
meta_df = pd.json_normalize(df["metadata_json"].apply(json.loads))
df = pd.concat([df.drop(columns=["metadata_json"]), meta_df], axis=1)

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

print(df.shape)                      # (391, ...)
print(df["source_name"].value_counts())
print(df[["question", "answer", "generation_sub_domain"]].head())

11.3 Using Hugging Face datasets

from datasets import load_dataset

ds = load_dataset("json", data_files="merged_finance_faq_serial.jsonl", split="train")
print(ds)
print(ds[0]["conversations"])

# Filter to just one institution, e.g. Siddhartha Bank
siddhartha_only = ds.filter(lambda row: row["source_name"] == "siddhartha_bank_faq")
print(len(siddhartha_only))   # 81

11.4 Converting to a simple prompt/completion fine-tuning format

import json

def to_prompt_completion(record):
    q = record["conversations"][0]["value"]
    a = record["conversations"][1]["value"]
    return {"prompt": q, "completion": a}

with open("merged_finance_faq_serial.jsonl", encoding="utf-8") as f_in, \
     open("finance_faq_prompt_completion.jsonl", "w", encoding="utf-8") as f_out:
    for line in f_in:
        record = json.loads(line)
        pc = to_prompt_completion(record)
        f_out.write(json.dumps(pc, ensure_ascii=False) + "\n")

---

12. Recommended Use Cases

• Instruction fine-tuning for Nepali-language finance assistants / chatbots.
• RAG knowledge base: index each answer (with its question and source_repo as metadata) for retrieval-augmented Nepali financial Q&A.
• Evaluation set: use held-out rows as a factual-accuracy benchmark for Nepali LLMs on finance topics.
• Domain-adaptive pretraining / continued pretraining on Nepali financial text.

13. Limitations & Caveats

• Domain skew: over half the dataset (204/391 rows, 52%) falls under the "Banking services" sub-domain, and just 3 sources (Siddhartha Bank, Prime Commercial Bank, NISM) account for ~49% of all rows — models trained on this data alone will be biased toward retail-banking FAQs.
• No multi-turn dialogue: every record is a single, isolated Q&A pair; there is no conversational history or follow-up-question handling represented.
• No negative/refusal examples: all answers are direct, informative responses; the dataset contains no examples of declining to answer, expressing uncertainty, or handling out-of-scope questions.
• url field is empty for all rows — original source page links were not retained during the merge.
• Nepali-only: english_content_allowed is false for every row, so no code-switched or bilingual content is present, despite some technical/brand terms appearing transliterated (e.g. डिपोजिटरी, म्युचुअल फन्ड).
• Static/institutional content: content reflects official published FAQ pages at the time of collection and may not capture the very latest product terms, rates, or fees (e.g. interest rates, minimum balances) from each institution.