sabin1234/Nepal_Finance_FAQ_Dataset

Merged Nepal Finance FAQ Dataset (Nepali)

A Nepali-language (Devanagari script) instruction-following dataset built from Frequently Asked Questions (FAQs) published by major Nepali financial institutions and fintech platforms. Each record is a single-turn human → gpt conversation extracted and normalized from real, publicly published FAQ pages, covering topics such as central banking, digital wallets, QR payments, letters of credit, mutual funds, and stock-market trading systems.

• File name: merged_nepal_finance_faq_ne.jsonl
• Format: JSON Lines (.jsonl) — one JSON object per line
• Total records: 818
• Language: Nepali (ne / ISO 639-3 npi), Devanagari script
• License: Apache-2.0 (permissive)
• Task type: Instruction-following (single-turn Q&A)
• Generation type: Real (sourced from genuine, published FAQ content — not synthetically generated)

---

---

1. Dataset Summary

Property	Value
Total records (rows)	818
Total unique sources (institutions)	8
Total unique FAQ categories	8
Total unique sub-domains	6
Domain (all records)	वित्त (Finance)
Conversation turns per record	2 (1 human + 1 gpt), 100% of records
Language	Nepali (ne / npi)
Script	Devanagari (Deva)
License	Apache-2.0
License tier	Permissive
Task type	instruction-following
Generation type	real (100%)
Condition	real (100%)
English content allowed	False (100%)
Duplicate IDs	0 (all 818 IDs are unique)
Empty/missing url field	818 (100% — no source URL recorded)
source_config	default (100%)
source_split	train (100%)
source_revision	v1 (100%)

---

2. File / Record Schema

Each line of the .jsonl file is a single JSON object with the following top-level keys:

Field	Type	Description	Example
id	string	Unique record identifier	"faq_000001"
conversations	array of objects	The single-turn dialogue. Each item has from ("human" or "gpt") and value (text)	see below
source	string	Human-readable name of the institution the FAQ came from	"eSewa"
source_name	string	Machine-readable slug identifying the specific FAQ collection	"esewa_taxpayer_incentive_prize_program_faq"
source_repo	string	Repository / organization the data was collected from (matches source in this dataset)	"eSewa"
source_config	string	Dataset configuration name (constant)	"default"
source_split	string	Original data split (constant)	"train"
source_revision	string	Version tag of the source snapshot (constant)	"v1"
source_row_id	string	Row ID in the original source (identical to id in every record)	"faq_000001"
language	string	ISO 639-1 language code	"ne"
language_code	string	ISO 639-3 language code	"npi"
script	string	Script code (ISO 15924)	"Deva"
license	string	License of the underlying content	"Apache-2.0"
license_tier	string	License category	"permissive"
task_type	string	NLP task category	"instruction-following"
generation_type	string	Whether content is real or synthetic	"real"
condition	string	Same as generation_type in this dataset	"real"
url	string	Source URL (empty for every record in this file)	""
metadata_json	string (stringified JSON)	Extra descriptive metadata — see Section 3	"{...}"

conversations sub-schema

"conversations": [
  { "from": "human", "value": "फोनपे क्यूआर के हो?" },
  { "from": "gpt",   "value": "फोनपे क्यूआर एक डिजिटल भुक्तानी समाधान हो ..." }
]

Sub-field	Type	Description
from	string	Speaker role — always either "human" (the question) or "gpt" (the answer)
value	string	The actual Nepali text content

Every single record in this dataset has exactly 2 turns: one human turn followed by one gpt turn (a classic FAQ question → answer pair, not a multi-turn chat).

---

3. Metadata Field (metadata_json) Schema

The metadata_json field is a stringified JSON object (you must json.loads() it separately) with the following keys:

Key	Type	Description	Example
generation_domain	string	Top-level domain	"वित्त" (Finance)
generation_category	string	The specific FAQ collection/category this record belongs to	"फोनपे सम्बन्धी प्रायः सोधिने प्रश्न"
generation_sub_domain	string	A finer-grained topical grouping	"डिजिटल भुक्तानी" (Digital Payments)
behavior	string	The intended response behavior/style label for the answer	"जानकारीमूलक उत्तर दिने" (Give an informative answer)
behavior_definition	string	A one-line definition explaining what the behavior label means	"मूल प्रश्नको अर्थ र मूल उत्तरको तथ्य कायम राख्दै नेपालीमा रूपान्तरण गरिएको उत्तर दिने।"
question_type	string	The category of question being asked	"तथ्यमा आधारित प्रश्न" (Fact-based question)
content_language	string	Language of the content, in Nepali	"नेपाली"
content_script	string	Script of the content, in Nepali	"देवनागरी"
english_content_allowed	boolean	Whether English text is permitted in the answer	false

---

4. Source Diversity

The dataset merges FAQ content from 8 distinct Nepali financial institutions/platforms, spanning central banking, commercial banking, digital wallets, QR/interbank payment infrastructure, the stock exchange, and mutual funds.

Source (institution)	source_name (slug)	Record count	% of dataset	Sector
Nepal Rastra Bank	nepal_rastra_bank_faq	388	47.4%	Central Bank / Monetary Policy
Khalti	khalti_official_faq	105	12.8%	Digital Wallet
Nepal Bank Limited	nepal_bank_limited_faq	105	12.8%	Commercial Banking
Fonepay	fonepay_faq	94	11.5%	QR / Digital Payments
Nepal Clearing House Limited (NCHL)	nchl_faq	89	10.9%	Payment Clearing System
Nepal Stock Exchange Limited (NEPSE)	nepse_faq	14	1.7%	Stock Market / Trading
eSewa	esewa_taxpayer_incentive_prize_program_faq	12	1.5%	Digital Payments (Tax Incentive Program)
Nabil Invest	nabil_invest_mutual_fund_faq	11	1.3%	Mutual Fund / Asset Management
Total		818	100%

Note: source, source_name, and source_repo map one-to-one for every record — one FAQ collection per institution, with no institution split across multiple categories.

---

5. Domain & Sub-domain Diversity

All 818 records share the same top-level generation_domain (वित्त / Finance), but are diversified across 8 categories and 6 sub-domains:

Categories (generation_category)

Category (Nepali)	English meaning	Record count
नेपाल राष्ट्र बैंक सम्बन्धी प्रायः सोधिने प्रश्न	Nepal Rastra Bank FAQs	388
खाल्टी सम्बन्धी प्रायः सोधिने प्रश्न	Khalti FAQs	105
नेपाल बैंक लिमिटेड सम्बन्धी प्रायः सोधिने प्रश्न	Nepal Bank Limited FAQs	105
फोनपे सम्बन्धी प्रायः सोधिने प्रश्न	Fonepay FAQs	94
एनसीएचएल सम्बन्धी प्रायः सोधिने प्रश्न	NCHL FAQs	89
नेप्से सम्बन्धी प्रायः सोधिने प्रश्न	NEPSE FAQs	14
करदाता प्रोत्साहन उपहार कार्यक्रम सम्बन्धी प्रायः सोधिने प्रश्न	Taxpayer Incentive Prize Program FAQs	12
नाबिल म्युचुअल फण्ड सम्बन्धी प्रायः सोधिने प्रश्न	Nabil Mutual Fund FAQs	11

Sub-domains (generation_sub_domain)

Sub-domain (Nepali)	English meaning	Record count	% of dataset
केन्द्रीय बैंक र मौद्रिक नीति	Central Bank & Monetary Policy	388	47.4%
डिजिटल भुक्तानी	Digital Payments	211	25.8%
बैंकिङ सेवा	Banking Services	105	12.8%
भुक्तानी प्रणाली	Payment Systems	89	10.9%
शेयर बजार र ट्रेड व्यवस्थापन प्रणाली	Stock Market & Trade Management System	14	1.7%
म्युचुअल फण्ड	Mutual Fund	11	1.3%
Total		818	100%

---

6. Behavior Diversity

The behavior and behavior_definition metadata fields describe the response style the gpt answer is meant to follow.

In this particular file, behavior diversity is intentionally uniform — all 818 records use a single behavior label. This reflects that the dataset was generated purely as a faithful Nepali FAQ transformation task (question in, factual answer out), rather than a multi-behavior instruction-tuning set (e.g. refusal, roleplay, summarization, etc.).

behavior (Nepali)	English meaning	behavior_definition (Nepali)	English meaning	Record count	% of dataset
जानकारीमूलक उत्तर दिने	Give an informative answer	मूल प्रश्नको अर्थ र मूल उत्तरको तथ्य कायम राख्दै नेपालीमा रूपान्तरण गरिएको उत्तर दिने।	Provide an answer translated/adapted into Nepali while preserving the meaning of the original question and the factual content of the original answer.	818	100%

If you plan to use this dataset alongside other behavior-diverse instruction datasets (e.g., ones containing refusals, clarifying questions, or multi-turn negotiation), be aware this file will only contribute the single "informative answer" behavior class and will not add behavioral diversity on its own.

---

7. Question Pattern Diversity

Although question_type metadata is uniform (तथ्यमा आधारित प्रश्न — "fact-based question" — for all 818 records), the surface-level phrasing/pattern of the human questions is diverse. Below is a breakdown of the linguistic question patterns found by scanning the human turn text for common Nepali interrogative markers. (Note: a single question can match more than one pattern, so percentages do not sum to 100%.)

Question pattern (Nepali marker)	English meaning	Matching records	% of dataset	Example question
... छ? (sentence-final "chha?")	Yes/No or state-confirmation question	368	45.0%	"एनएफसीएफका युनिट कसरी किन्न सकिन्छ?" (pattern also appears mid-sentence)
सक्छ / सकिन्छ ...?	"Can it be done? / Is it possible?" (capability question)	149	18.2%	"फोनपे क्यूआर प्रयोग गर्न सुरक्षित छ?"
कसरी	"How"	133	16.3%	"फोनपे क्यूआर कसरी प्रयोग गर्न सकिन्छ?"
के हो	"What is" (definition question)	116	14.2%	"फोनपे क्यूआर के हो?"
कति	"How much / how many"	94	11.5%	—
किन	"Why"	88	10.8%	—
के गर्ने / के गर्न सकिन्छ	"What should be done if..." (troubleshooting question)	30	3.7%	"कारोबार व्यवस्थापन प्रणालीमा दोहोरो बीओआईडी त्रुटि आएमा के गर्ने?"
कसले	"Who"	20	2.4%	"एनसीएचएल-ईसीसी कसले नियमन गर्छ?"
फाइदा / लाभ	"What are the benefits/advantages"	13	1.6%	"पार्टनरको रूपमा खाल्टी बजार प्रयोग गर्दा के फाइदा छन्?"
कहिले	"When"	6	0.7%	—

Question ending punctuation style

Ending style	Record count	% of dataset
Ends with ? (question mark)	814	99.5%
Ends with : (colon, keyword-style prompt)	2	0.24%
Ends with । (Devanagari full stop)	1	0.12%
Other / no terminal punctuation	1	0.12%

The two colon-terminated and one full-stop-terminated entries come from the eSewa source and are short topical labels (e.g. "विद्युतीय भुक्तानी:" — "Electronic Payment:") rather than full interrogative sentences — these act as FAQ headings that the corresponding answer expands on.

Question intent categories (qualitative grouping)

Beyond surface markers, the human questions can be grouped into four broad intents:

Intent category	Description	Approx. share
Definitional ("What is X?")	Asks for a definition or explanation of a term, product, or service	~14%
Procedural ("How do I do X?")	Asks for step-by-step instructions or a process	~16%
Eligibility / capability ("Can I / who can...?")	Asks whether an action is possible or who is authorized/responsible	~18–20%
Troubleshooting / conditional ("What to do if X happens?")	Describes an error/edge case and asks for the resolution	~4%

---

8. Conversation Structure Statistics

Property	Value
Records with exactly 2 turns (1 human + 1 gpt)	818 (100%)
Records with role order human → gpt	818 (100%)
Records with any other turn count or role order	0
Duplicate id values	0
id == source_row_id for all records	Yes (100%)

---

9. Text Length Statistics

Character-length statistics computed over the raw value strings (Devanagari text, so character count roughly tracks glyph count, not Latin "words").

Turn	Min length (chars)	Max length (chars)	Average length (chars)
human (question)	11	777	58.8
gpt (answer)	27	1,507	247.7

Answers are, on average, roughly 4.2x longer than questions — consistent with FAQ answers that include step-by-step instructions, lists of required documents, or multi-clause explanations.

---

10. Sample Records

One example per source, showing the raw structure:

eSewa

{
  "id": "faq_000001",
  "conversations": [
    {"from": "human", "value": "विद्युतीय भुक्तानी:"},
    {"from": "gpt", "value": "विद्युतीय भुक्तानी योग्य कारोबार स्वतः कार्यक्रममा समावेश हुनेछन्।"}
  ],
  "source": "eSewa",
  "source_name": "esewa_taxpayer_incentive_prize_program_faq"
}

Fonepay

{
  "conversations": [
    {"from": "human", "value": "फोनपे क्यूआर के हो?"},
    {"from": "gpt", "value": "फोनपे क्यूआर एक डिजिटल भुक्तानी समाधान हो जसले क्विक रेस्पोन्स (क्यूआर) कोड प्रयोग गर्दछ..."}
  ],
  "source": "Fonepay"
}

Nepal Rastra Bank

{
  "conversations": [
    {"from": "human", "value": "नेपाल राष्ट्र बैंकको दूरदृष्टि के रहेको छ?"},
    {"from": "gpt", "value": "नेपाल राष्ट्र बैंकको दूरदृष्टि “आधुनिक, गतिशील, विश्वसनीय र प्रभावकारी केन्द्रीय बैंक” भन्ने रहेको छ।"}
  ],
  "source": "Nepal Rastra Bank"
}

For all other sources (Khalti, Nepal Bank Limited, NCHL, NEPSE, Nabil Invest), see Section 4 for slugs you can filter on using the code in Section 11.

---

11. How to Load the Dataset

11.1 Plain Python (standard library only)

import json

path = "merged_nepal_finance_faq_ne.jsonl"

records = []
with open(path, "r", encoding="utf-8") as f:
    for line in f:
        line = line.strip()
        if line:
            records.append(json.loads(line))

print(f"Loaded {len(records)} records")
print(records[0])

# Parse the nested metadata_json string into a dict
for r in records:
    r["metadata"] = json.loads(r["metadata_json"])

# Access first record's question, answer, and behavior
first = records[0]
question = first["conversations"][0]["value"]
answer   = first["conversations"][1]["value"]
behavior = first["metadata"]["behavior"]

print("Q:", question)
print("A:", answer)
print("Behavior:", behavior)

11.2 Using pandas

import json
import pandas as pd

path = "merged_nepal_finance_faq_ne.jsonl"

# jsonlines can be read directly with lines=True
df = pd.read_json(path, lines=True)

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

# Extract question/answer text from the conversations column
df["question"] = df["conversations"].apply(lambda c: c[0]["value"])
df["answer"]   = df["conversations"].apply(lambda c: c[1]["value"])

print(df.shape)
print(df[["id", "source", "generation_sub_domain", "question", "answer"]].head())

# Example: filter records from Nepal Rastra Bank only
nrb_df = df[df["source"] == "Nepal Rastra Bank"]
print(f"Nepal Rastra Bank records: {len(nrb_df)}")

# Example: group by source and count
print(df.groupby("source").size().sort_values(ascending=False))

11.3 Using the Hugging Face datasets library

from datasets import load_dataset

# Load directly from a local jsonl file
dataset = load_dataset(
    "json",
    data_files="merged_nepal_finance_faq_ne.jsonl",
    split="train",
)

print(dataset)
print(dataset[0])

# Filter by source using .filter()
khalti_only = dataset.filter(lambda x: x["source"] == "Khalti")
print(f"Khalti records: {len(khalti_only)}")

# Map to flatten metadata_json for easier downstream use
import json

def flatten_metadata(example):
    meta = json.loads(example["metadata_json"])
    example.update(meta)
    return example

dataset = dataset.map(flatten_metadata)
print(dataset.column_names)

11.4 Using jsonlines package (streaming, memory-efficient)

import jsonlines

path = "merged_nepal_finance_faq_ne.jsonl"

with jsonlines.open(path) as reader:
    for obj in reader:
        q = obj["conversations"][0]["value"]
        a = obj["conversations"][1]["value"]
        # process one record at a time without loading the whole file into memory
        print(obj["id"], "-", q[:50])

---

12. Quick Data Exploration Recipes

import json
from collections import Counter

path = "merged_nepal_finance_faq_ne.jsonl"
records = [json.loads(l) for l in open(path, encoding="utf-8") if l.strip()]
for r in records:
    r["metadata"] = json.loads(r["metadata_json"])

# 1) Count records per source
print(Counter(r["source"] for r in records))

# 2) Count records per sub-domain
print(Counter(r["metadata"]["generation_sub_domain"] for r in records))

# 3) Find the longest answer in the dataset
longest = max(records, key=lambda r: len(r["conversations"][1]["value"]))
print("Longest answer source:", longest["source"])
print("Longest answer length:", len(longest["conversations"][1]["value"]))

# 4) Find all questions containing "कसरी" (how)
how_questions = [
    r["conversations"][0]["value"]
    for r in records
    if "कसरी" in r["conversations"][0]["value"]
]
print(f"Found {len(how_questions)} 'how' questions")

# 5) Build a simple instruction-tuning formatted list (e.g. for fine-tuning)
formatted = [
    {"instruction": r["conversations"][0]["value"],
     "output": r["conversations"][1]["value"]}
    for r in records
]
print(formatted[0])

---

13. Data Quality Notes / Known Limitations

Note	Detail
Single behavior class	Every record shares one behavior label (जानकारीमूलक उत्तर दिने) — no refusal, clarification, or multi-behavior diversity exists in this file.
Single question-type label	The metadata question_type field is constant across all records (तथ्यमा आधारित प्रश्न), even though surface question patterns vary (see Section 7).
No source URLs	The url field is empty ("") for all 818 records — original page links were not retained.
Imbalanced source distribution	Nepal Rastra Bank alone accounts for ~47% of the dataset; smaller sources like Nabil Invest (11 records) and eSewa (12 records) are comparatively under-represented. Consider stratified sampling if class balance matters for your use case.
Nepali-only, no code-switching	english_content_allowed is False for all records — content is Devanagari-only, so this dataset is not suited for code-switched (Nepali-English) evaluation.
Single-turn only	There are no multi-turn conversations — every record is exactly one Q→A pair, so this dataset alone cannot be used for multi-turn dialogue training.
Fixed split/config/revision metadata	source_config, source_split, and source_revision are constant (default, train, v1) — these fields carry no discriminative information in this file.

---

14. Suggested Use Cases

• Fine-tuning or instruction-tuning Nepali-language LLMs on finance-domain Q&A.
• Building a retrieval-augmented generation (RAG) knowledge base for Nepali fintech/banking customer support.
• Evaluating Nepali NLP models on domain-specific factual question answering.
• Building a finance-domain chatbot or FAQ search tool for Nepali banking/fintech platforms.
• As one of several source datasets in a larger, more behavior-diverse Nepali instruction-tuning mixture.

---

15. License & Attribution

• License: Apache-2.0 (permissive) — see the license and license_tier fields in every record.
• Content origin: Real, publicly published FAQ content from Nepal Rastra Bank, Khalti, Nepal Bank Limited, Fonepay, Nepal Clearing House Limited (NCHL), Nepal Stock Exchange Limited (NEPSE), eSewa, and Nabil Invest.
• Please retain attribution to the original institutions when redistributing derived answers, and verify current terms directly with each institution if you plan on commercial redistribution, since FAQ content may be updated or changed by the source institutions over time.