sabin1234/Nepal_Banking_FAQ

Merged Nepali FAQ Dataset (merged_faq_nepali_dataset)

A merged, instruction-tuning-ready collection of real-world Nepali-language FAQ pairs drawn from banking, cooperative, insurance, pension, and digital-payment institutions in Nepal and India. Each record is a single-turn human → gpt conversation, formatted in the common ShareGPT / conversations JSONL schema used for supervised fine-tuning (SFT) of instruction-following LLMs.

---

1. Dataset Summary

Property	Value
File	merged_faq_nepali_dataset__3_.jsonl
Format	JSON Lines (one JSON object per line)
Total records	308
Conversation turns per record	2 (human then gpt) — 100% single-turn
Language	Nepali (ne / ISO 639-3 npi)
Script	Devanagari (Deva)
License	Apache-2.0 (permissive)
Task type	instruction-following
Generation type	real (sourced from genuine institutional FAQ pages, not synthetically generated)
Distinct source organizations	9
Distinct sub-domains / FAQ categories	22
Duplicate IDs	None — all 308 id values are unique

This is a small, domain-specific dataset, best suited for fine-tuning or evaluating a Nepali-language financial-services assistant, for benchmarking Nepali NLU/NLG models on factual QA, or as a seed set to bootstrap a larger synthetic FAQ corpus.

---

2. Where the Data Comes From (Source Distribution)

The dataset merges FAQ content scraped/curated from 9 real institutions, each represented by a source_name (dataset shard identifier). Distribution of records per source:

Source Institution	source_name	Records	% of Total
Department of Financial Services, Government of India	dfsindia_faq	90	29.2%
Everest Bank (Nepal)	everest_bank_faq	69	22.4%
कर्मचारी सञ्चय कोष — Employees Provident Fund	employees_provident_fund_faq	43	14.0%
Agricultural Development Bank Limited (ADBL) — Online Account	adbl_online_account_faq	19	6.2%
Agricultural Development Bank Limited (ADBL) — Kisan App	adbl_kisan_faq	15	4.9%
COPOMIS (Cooperative Management Information System)	copomis_faq	29	9.4%
ईसेवा — eSewa (digital wallet)	esewa_faq	18	5.8%
Citizen Life Insurance Company Limited	citizenlife_faq	16	5.2%
Deposit & Credit Guarantee Fund (DCGF)	dcgf_faq	6	1.9%
CDS and Clearing Limited (CDSC)	cdsc_faq	3	1.0%

Note: adbl_online_account_faq and adbl_kisan_faq are both grouped under the parent source / source_repo value "Agricultural Development Bank Limited" (34 records combined), since source_name is the finer-grained shard identifier while source / source_repo identify the parent institution.

Broader Domain Grouping (generation_domain)

Domain	Records	%
वित्तीय सेवा (Financial Services — pensions, NPS, APY)	202	65.6%
बैंकिङ (Banking)	40	13.0%
सहकारी (Cooperatives)	29	9.4%
डिजिटल भुक्तानी (Digital Payments)	18	5.8%
बीमा (Insurance)	16	5.2%
बैंकिङ तथा पूँजी बजार (Banking & Capital Markets)	3	1.0%

Fine-Grained Sub-Domain / FAQ Category (generation_sub_domain)

Sub-domain	Records
राष्ट्रिय पेन्सन प्रणाली (National Pension System)	47
अटल पेन्सन योजना (Atal Pension Yojana)	43
सहकारी व्यवस्थापन सूचना प्रणाली (COPOMIS)	29
खाता व्यवस्थापन तथा इ-सेवा (Account Mgmt & e-Services)	21
डेबिट कार्ड तथा सामान्य सेवा (Debit Card & General Services)	21
एडीबीएल अनलाइन खाता (ADBL Online Account)	19
डिजिटल वालेट (Digital Wallet — eSewa)	18
सिटीजन लाइफ इन्स्योरेन्स (Citizen Life Insurance)	16
एडीबीएल किसान एप (ADBL Kisan App)	15
सञ्चयकर्ता स्वास्थोपचार योजना (Depositor Health Scheme)	13
इन्टरनेट बैंकिङ (Internet Banking)	10
रेमिटेन्स (Remittance)	10
कार्ड व्यवस्थापन तथा शुल्क (Card Management & Fees)	9
डिम्याट शेयर तथा धितोपत्र (Demat Shares & Securities)	8
निक्षेप तथा कर्जा सुरक्षण कोष (Deposit & Credit Guarantee)	6
स्विफ्ट तथा अन्तर्राष्ट्रिय स्थानान्तरण (SWIFT & Int'l Transfer)	6
सामान्य सेवा (General Services)	5
सीडीएस र क्लियरिङ (CDS & Clearing)	3
सापटी तथा कर्जा (Loans & Credit)	3
मिस्ड कल अलर्ट (Missed Call Alert)	3
सम्पर्क तथा प्रतिनिधि कार्यालय (Contact & Branch Offices)	2
प्रसूति स्याहार तथा परिवार (Maternity & Family Care)	1

---

3. Language / Content Composition

• 100% Nepali-script (Devanagari) content. content_language = "नेपाली" and content_script = "देवनागरी" for all 308 rows; english_content_allowed is false for every record.
• English tokens still appear in 9 of 308 records (~2.9%) — but only as unavoidable proper nouns / contact info embedded inside otherwise-Nepali sentences, e.g. email addresses (sahakaribivag@gmail.com), URLs (www.everestbankltd.com), app names (Kisan App), or a SWIFT code (EVBLNPKA). No record is English-medium; these are just inline artifacts (URLs, emails, brand/app names) inside Nepali sentences.
• Answer text is factual/institutional in register — formal Nepali, often containing numbers, step sequences (using "=" as a step separator, e.g. वेबसाइटमा जानुहोस् = खाता खोल्नुहोस् = फारम भर्नुहोस्), or short direct answers ("हो," / "होइन,").

---

4. Behaviour Distribution

The behavior field in metadata_json is uniform across the entire dataset — every single record was generated/curated with the same target model behaviour:

behavior	Records	%
जानकारीमूलक उत्तर दिने (Give an informative/factual answer)	308	100%

However, the behavior_definition (the precise instruction the annotator/generator followed) has two variants, differing only by whether "शुद्ध" (pure/standard) Nepali is emphasized:

behavior_definition (English gloss)	Records	%
"Preserve the meaning of the original question and the facts of the original answer, translated into pure (शुद्ध) Nepali."	255	82.8%
"Preserve the meaning of the original question and the facts of the original answer, translated into Nepali."	53	17.2%

Interpretation: This dataset was built by taking source FAQ Q&A pairs (some originally in English, e.g. the Indian government's DFS/APY/NPS FAQs) and translating/localizing them into Nepali while strictly preserving factual content — i.e., this is a faithful-translation + informative-answering behavior dataset, not a creative-generation or multi-behavior dataset. There is no variation in behavior category itself (no "refuse", "clarify", "small talk", "opinion" behaviors present) — 100% of examples are single-behavior factual QA.

---

5. Question Type / Question Pattern Distribution

The question_type metadata field is likewise constant:

question_type	Records	%
तथ्यमा आधारित प्रश्न (Fact-based question)	308	100%

Since question_type doesn't vary, deeper pattern insight comes from analyzing the actual interrogative structure of the human turn (the Nepali question word / pattern used). Classifying by first-matching interrogative marker in each question:

Nepali Question Marker	English meaning	Records	%
के	"what / is it (yes-no)"	75	24.4%
कसरी	"how"	62	20.1%
(other — no simple marker, e.g. definitional/imperative phrasing)	—	40	13.0%
किन	"why"	38	12.3%
के हो	"what is..."	28	9.1%
कति	"how much / how many"	24	7.8%
कुन	"which"	17	5.5%
के गर्ने	"what should be done"	9	2.9%
कहिले	"when"	9	2.9%
कसले	"who"	3	1.0%
कहाँ	"where"	3	1.0%

Takeaways:

• The dataset is dominated by definitional/eligibility/yes-no questions (के…) and procedural how-to questions (कसरी…) — together ~45% of all questions. This matches the institutional-FAQ nature of the source (eligibility criteria, account-opening steps).
• "किन" (why) and "कति" (how much/many) questions cover reasoning and numeric/fee/limit queries respectively — common in banking/pension FAQs (e.g. interest rates, contribution amounts, age limits).
• Very few "कहाँ" (where) and "कसले" (who) questions — the dataset is weak on location/branch-office and stakeholder-identification queries, reflected also by the low record count (2) for the "सम्पर्क तथा प्रतिनिधि कार्यालय" sub-domain above.

---

6. Text Length Statistics

Field	Min (chars)	Max (chars)	Avg (chars)
Question (human turn)	13	135	~52.5
Answer (gpt turn)	5	716	~196.4

Questions are short and single-sentence; answers range from very terse ("हो," / "होइन,") to fairly long multi-step procedural explanations (max 716 characters), reflecting the mix of yes/no eligibility answers vs. detailed how-to instructions.

---

7. Full Schema Reference

Each JSONL line is a JSON object with the following top-level fields:

Field	Type	Description
id	string	Unique record identifier (e.g. adbl_faq_00001). Unique across the whole file.
conversations	array[object]	Exactly 2 turns: {"from": "human", "value": "<question>"}, {"from": "gpt", "value": "<answer>"}.
source	string	Human-readable name of the originating institution.
source_name	string	Machine-readable shard/sub-source identifier (finer-grained than source).
source_repo	string	Repository/collection name the record was pulled from (usually same as source).
source_config	string	Dataset config name — always "default".
source_split	string	Original split — always "train".
source_revision	string	Source dataset version tag — always "v1".
source_row_id	string	Row ID within the original source shard (may differ from top-level id after merge — see note below).
language	string	ISO 639-1 code — always "ne".
language_code	string	ISO 639-3 code — always "npi".
script	string	ISO 15924 script code — always "Deva" (Devanagari).
license	string	Always "Apache-2.0".
license_tier	string	Always "permissive".
task_type	string	Always "instruction-following".
generation_type	string	Always "real" (not model-synthesized).
condition	string	Always "real".
url	string	Source URL — empty ("") for every record in this file.
metadata_json	string (stringified JSON)	Nested metadata — must be parsed separately with json.loads(). See below.

metadata_json (nested, stringified) fields:

Field	Description
generation_domain	Broad domain in Nepali (e.g. "बैंकिङ", "वित्तीय सेवा").
generation_category	Human-readable FAQ category/page title in Nepali.
generation_sub_domain	Fine-grained product/scheme name in Nepali.
behavior	Target model behavior label — always "जानकारीमूलक उत्तर दिने".
behavior_definition	Full instruction describing the translation/answering behavior expected.
question_type	Always "तथ्यमा आधारित प्रश्न" (fact-based question).
content_language	Always "नेपाली".
content_script	Always "देवनागरी".
english_content_allowed	Boolean — always false.

Note on source_row_id vs id: in some records (e.g. dfsindia_000101) the top-level id and the source_row_id differ (dfsindia_000101 vs dfsindia_000013), indicating the merge/dedup process re-indexed some rows while preserving the original row's source-side ID for traceability.

---

8. How to Load the Dataset

8.1 Plain Python (no dependencies)

import json

path = "merged_faq_nepali_dataset__3_.jsonl"

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

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

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

# Example: print the first Q&A pair with its category
first = records[0]
question = first["conversations"][0]["value"]
answer   = first["conversations"][1]["value"]
category = first["metadata"]["generation_category"]

print("Category:", category)
print("Q:", question)
print("A:", answer)

8.2 Using pandas

import json
import pandas as pd

path = "merged_faq_nepali_dataset__3_.jsonl"

# jsonlines -> DataFrame
df = pd.read_json(path, lines=True)

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

# Expand metadata_json (stringified JSON) 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)

print(df.shape)
print(df[["id", "source_name", "question", "answer", "generation_domain"]].head())

# Example: distribution of records per source
print(df["source_name"].value_counts())

8.3 Using Hugging Face datasets

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files="merged_faq_nepali_dataset__3_.jsonl",
    split="train",
)

print(ds)
print(ds[0])

# Parse metadata_json on the fly with .map()
import json
def expand_metadata(example):
    meta = json.loads(example["metadata_json"])
    example.update(meta)
    return example

ds = ds.map(expand_metadata)
print(ds.column_names)

8.4 Formatting for SFT / chat-template fine-tuning

Since the schema already uses the conversations (ShareGPT-style) format, it maps directly to most fine-tuning frameworks (Axolotl, LLaMA-Factory, TRL, etc.) with little to no transformation. Example conversion to OpenAI/ChatML-style messages for use with transformers chat templates:

import json

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

with open("merged_faq_nepali_dataset__3_.jsonl", encoding="utf-8") as f:
    records = [json.loads(l) for l in f if l.strip()]

chatml_records = [to_chatml(r) for r in records]

with open("chatml_format.jsonl", "w", encoding="utf-8") as f:
    for rec in chatml_records:
        f.write(json.dumps(rec, ensure_ascii=False) + "\n")

---

9. Suggested Uses

• Fine-tuning / instruction-tuning small Nepali-capable LLMs for financial and government-service customer support.
• Evaluation benchmark for Nepali factual QA and eligibility/how-to reasoning.
• Seed set for data augmentation — e.g. prompting a larger LLM to generate paraphrased variants of each Q&A pair while preserving the same behavior_definition (pure-Nepali, fact-preserving translation).
• Domain-adaptation corpus for retrieval-augmented generation (RAG) systems serving Nepali fintech users (build a vector index over the answer field, keyed by generation_sub_domain).

10. Known Limitations

• Small size (308 records) — not sufficient alone for full fine-tuning of a general-purpose model; best combined with other Nepali instruction data.
• Single-turn only — no multi-turn dialogue, follow-up questions, or clarification behavior represented.
• Single behavior class — entirely "informative answer" behavior; no examples of refusal, clarification-seeking, or opinion/subjective responses.
• Domain-skewed — pension schemes (NPS + APY) alone account for ~29% of the dataset; some sub-domains (e.g. maternity care, missed-call alerts) have only 1–3 examples.
• Residual English artifacts — ~2.9% of records contain inline English proper nouns (URLs, emails, SWIFT codes) despite english_content_allowed: false, worth normalizing/masking if strict Nepali-only text is required.
• No url provenance — the url field is empty for all records, so original source pages cannot be traced directly from the file alone.