sabin1234/ShareHub_FAQ_Nepali_Dataset

ShareHub FAQ Nepali Dataset

A Nepali-language, instruction-following (question–answer) dataset of 2,448 real frequently-asked-questions about securities listed on the Nepal Stock Exchange (NEPSE) — covering common stocks, debentures, and mutual funds — sourced from ShareHub. Every record is a single-turn human → gpt conversation in Devanagari script, and every record carries rich provenance and generation metadata.

File analyzed: sharehub_faq.jsonl

---

1. Quick Facts

Property	Value
File format	JSON Lines (.jsonl) — one JSON object per line
Total records (rows)	2,448
Total conversations	2,448 (each with exactly 2 turns: human + gpt)
Total turns	4,896
Language	Nepali (ne / ISO 639-3 npi)
Script	Devanagari (Deva)
License	Apache-2.0 (permissive tier)
Task type	Instruction-following
Generation type	real (real-world sourced, not synthetically fabricated)
Source	ShareHub (sharehub_faq)
Unique securities covered	415 distinct symbols (companies / instruments)
Distinct question patterns (templates)	8
Distinct security types	3 (stock, debenture, mutual_fund)
Distinct generation sub-domains (sectors)	15
Behavior label	1 constant behavior across the whole dataset (see §5)
Question type label	1 constant type across the whole dataset (see §5)
Average question length	81.0 characters
Average answer length	236.7 characters
English content allowed	false (answers are pure Nepali)

---

2. Record (Schema) Structure

Each line of the .jsonl file is one JSON object with the following top-level fields:

Field	Type	Description
id	string	Unique record identifier, pattern sharehub_<symbol-lowercase>_<3-digit-index>, e.g. sharehub_aclbsl_001
conversations	array of objects	The single-turn dialogue: a list of exactly 2 turn objects (human then gpt)
source	string	Always "ShareHub" — the originating platform/brand
source_name	string	Always "sharehub_faq" — the internal dataset name
source_repo	string	Always "ShareHub" — the source repository name
source_config	string	Always "default" — the dataset configuration used
source_split	string	Always "train" — the data split this record belongs to
source_revision	string	Always "v1" — the version/revision tag of the source
source_row_id	string	Identical to id — the original row identifier in the source dataset
language	string	Always "ne" — ISO 639-1 language code (Nepali)
language_code	string	Always "npi" — ISO 639-3 language code (Nepali)
script	string	Always "Deva" — ISO 15924 script code (Devanagari)
license	string	Always "Apache-2.0"
license_tier	string	Always "permissive"
task_type	string	Always "instruction-following"
generation_type	string	Always "real" — indicates the Q&A is grounded in real data, not hallucinated
condition	string	Always "real"
url	string	Always empty string "" in this dataset (no source URL attached)
metadata_json	string (JSON-encoded)	A nested JSON object (stored as a string) holding generation-specific metadata — see §3

2.1 conversations sub-structure

Each element of conversations is an object with two fields:

Field	Type	Description
from	string	Either "human" (the question) or "gpt" (the answer)
value	string	The actual text content of that turn, in Nepali (Devanagari)

Every conversation in this dataset has exactly 2 turns — one human question followed by one gpt answer. There are no multi-turn dialogues.

---

3. The metadata_json Field (Decoded)

metadata_json is a JSON object serialized as a string inside each row. Once parsed, it contains 11 keys:

Key	Type	Description
generation_domain	string	Broad domain of the question — always "वित्त" ("Finance")
generation_category	string	Category label — always "शेयर बजार सम्बन्धी प्रायः सोधिने प्रश्न" ("Frequently asked questions about the stock market")
generation_sub_domain	string	The specific sector/industry of the security being asked about (15 distinct values — see §5.3)
behavior	string	The response behavior the answer follows — always "जानकारीमूलक उत्तर दिने" ("Provide an informative answer")
behavior_definition	string	A full-sentence definition of the behavior (constant — see §5.1)
question_type	string	Category of question — always "तथ्यमा आधारित प्रश्न" ("Fact-based question")
content_language	string	Always "नेपाली" ("Nepali")
content_script	string	Always "देवनागरी" ("Devanagari")
english_content_allowed	boolean	Always false — no English is permitted in the generated answer
original_symbol	string	The NEPSE ticker symbol of the security this Q&A pair is about, e.g. ACLBSL, ADBL, ADBLB (415 distinct values)
security_type	string	Type of the financial instrument — one of stock, debenture, mutual_fund

Important: metadata_json must be parsed with a JSON decoder (e.g. json.loads()), it is not a native nested object in the raw line — it is a string containing escaped JSON.

---

4. Full Example Records (Decoded)

Example 1 — Stock / Microfinance

{
  "id": "sharehub_aclbsl_001",
  "conversations": [
    {
      "from": "human",
      "value": "Aarambha Chautari Laghubitta Bittiya Sanstha Limited (ACLBSL) को कुल बजार पूँजीकरण (मार्केट क्याप) र मूल्यांकन के हो?"
    },
    {
      "from": "gpt",
      "value": "Aarambha Chautari Laghubitta Bittiya Sanstha Limited (ACLBSL) को कुल बजार पूँजीकरण नेरू ३,२४,९२,१९,९७५ (३.२ अरब) र फ्लोट बजार पूँजीकरण नेरू १,२९,९६,८७,९९० (१.३ अरब) छ। बुक भ्याल्यू १८०.९ छ, जसले मूल्य-देखि-बुक मूल्य ४.८९ हुन्छ। प्रत्येक शेयरको फेस भ्याल्यू रु. १०० छ।"
    }
  ],
  "source": "ShareHub",
  "source_name": "sharehub_faq",
  "source_repo": "ShareHub",
  "source_config": "default",
  "source_split": "train",
  "source_revision": "v1",
  "source_row_id": "sharehub_aclbsl_001",
  "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": "शेयर बजार सम्बन्धी प्रायः सोधिने प्रश्न",
    "generation_sub_domain": "Microfinance",
    "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,
    "original_symbol": "ACLBSL",
    "security_type": "stock"
  }
}

Example 2 — Debenture / Non-Convertible Debenture

{
  "id": "sharehub_adblb_001",
  "conversations": [
    {
      "from": "human",
      "value": "4% Agricultural Bond (ADBLB) को कुल बजार पूँजीकरण (मार्केट क्याप) र मूल्यांकन के हो?"
    },
    {
      "from": "gpt",
      "value": "४% Agricultural Bond (ADBLB) को कुल बजार पूँजीकरण नेरू ६,००,००,००,००० (६ अरब) र फ्लोट बजार पूँजीकरण नेरू ० छ। बुक भ्याल्यू ० छ, जसले मूल्य-देखि-बुक मूल्य ० हुन्छ। प्रत्येक शेयरको फेस भ्याल्यू रु. १००० छ।"
    }
  ],
  "metadata_json": {
    "generation_sub_domain": "Non-Convertible Debenture",
    "original_symbol": "ADBLB",
    "security_type": "debenture",
    "...": "(other fields identical to Example 1)"
  }
}

(Only the differing / most relevant fields are shown for Example 2 for brevity; all other top-level fields follow the same constant values documented in §2.)

---

5. Diversity Analysis

Although several fields are constant across the entire dataset (because the dataset was generated under one fixed generation pipeline — one domain, one category, one behavior, one question type), the real diversity in this dataset lives in three places: (a) the behavior/question-type metadata labels (constant, documented below for completeness), (b) the 8 underlying question patterns / templates, and (c) the 15 sector sub-domains × 3 security types × 415 companies the questions are actually about. All are broken out in full below.

5.1 Behavior Field (Constant)

behavior value (Nepali)	English meaning	behavior_definition (Nepali)	English meaning	Row count
जानकारीमूलक उत्तर दिने	"Provide an informative answer"	मूल प्रश्नको अर्थ र मूल उत्तरको तथ्य कायम राख्दै नेपालीमा रूपान्तरण गरिएको उत्तर दिने।	"Give an answer translated into Nepali while preserving the meaning of the original question and the facts of the original answer."	2,448 (100%)

Note: This dataset was generated with a single fixed behavior mode — every answer is a faithful, fact-preserving, informational Nepali rendering of the underlying financial data point. There is no variation in behavior type (e.g., no "refusal," "clarifying question," "opinion," or "creative" behaviors present) — the entire corpus is uniformly informational/factual by design.

5.2 Question Type Field (Constant)

question_type value (Nepali)	English meaning	Row count
तथ्यमा आधारित प्रश्न	"Fact-based question"	2,448 (100%)

5.3 Question Pattern / Template Diversity (8 Templates)

While the question_type metadata label is uniform, the actual questions follow 8 distinct underlying templates, each asking about a different financial attribute of a security. The company/security name is substituted into each template. These were identified by stripping the company name + ticker prefix from every question and clustering identical remainders:

#	Question template (Nepali, as it appears after the company name)	English translation	Occurrences	% of dataset
1	...को कुल बजार पूँजीकरण (मार्केट क्याप) र मूल्यांकन के हो?	"What is [X]'s total market capitalization (market cap) and valuation?"	402	16.4%
2	...का कति शेयरहरू नेप्सेमा सूचीकृत छन्?	"How many shares of [X] are listed on NEPSE?"	415	17.0%
3	...को पछिल्लो कारोबार मूल्य कति छ?	"What is [X]'s latest traded price?"	402	16.4%
4	...को ऐतिहासिक मूल्य प्रवृत्ति र प्रदर्शन कस्तो छ?	"What has [X]'s historical price trend and performance been like?"	402	16.4%
5	...का मुख्य वित्तीय सूचकहरू र प्रदर्शन मापदण्डहरू के-के हुन्?	"What are [X]'s key financial indicators and performance metrics?"	401	16.4%
6	...को खुला दायरा कति हो?	"What is [X]'s open range (IPO/issue price range)?"	238	9.7%
7	...को प्रवर्द्धक अनलक मिति कहिले हो?	"When is [X]'s promoter-share unlock date?"	157	6.4%
8	...को परिपक्वता अवधि कति हो?	"What is [X]'s maturity period?" (debenture-specific)	31	1.3%

Total: 2,448 questions across 8 templates.

Templates 1–5 are asked for almost every security (near-universal coverage, ~400+ each). Template 6 (open range) applies mainly to debentures and mutual funds with IPO/issue ranges. Template 7 (promoter unlock date) applies to stocks with promoter lock-in periods. Template 8 (maturity period) applies only to debentures, which explains its low count (31, matching debenture-type maturity questions).

5.4 Security Type Diversity

security_type	English meaning	Row count	% of dataset
stock	Ordinary/common equity shares	1,702	69.5%
debenture	Fixed-income debt instrument (bond)	440	18.0%
mutual_fund	Collective investment scheme unit	306	12.5%

5.5 Sector / Sub-Domain Diversity (15 Sub-Domains)

generation_sub_domain describes the industry/sector of the underlying security. Cross-referenced with security_type:

#	Sub-domain (sector)	Security type	Row count	% of dataset
1	Hydro Power	stock	743	30.4%
2	Non-Convertible Debenture	debenture	440	18.0%
3	Mutual Fund	mutual_fund	295	12.1%
4	Microfinance	stock	270	11.0%
5	Manufacturing And Processing	stock	114	4.7%
6	Commercial Banks	stock	95	3.9%
7	Development Banks	stock	82	3.3%
8	Life Insurance	stock	81	3.3%
9	Finance	stock	75	3.1%
10	Non Life Insurance	stock	69	2.8%
11	Others	stock	62	2.5%
12	Hotels And Tourism	stock	52	2.1%
13	Investment	stock	47	1.9%
14	Tradings	stock	12	0.5%
15	Open Ended Mutual Fund (SIP)	mutual_fund	11	0.4%

Total: 2,448 rows across 15 sub-domains. Hydro Power is by far the largest single sector (30.4% of the entire dataset), reflecting NEPSE's heavy listing of hydropower companies.

5.6 Per-Security Question Coverage

Each of the 415 unique securities (identified by original_symbol) was asked a different number of the 8 question templates, depending on its instrument type (e.g., debentures get the maturity-period question; stocks get the promoter-unlock question; not every security has an "open range" recorded):

Questions asked per security	Number of securities	% of all 415 securities
7 questions	136	32.8%
6 questions	151	36.4%
5 questions	114	27.5%
4 questions	1	0.2%
3 questions	1	0.2%
2 questions	1	0.2%
1 question	11	2.7%

Most securities (96.7%) received between 5 and 7 questions, covering the core template set (market cap, listed shares, last traded price, historical trend, financial indicators) plus applicable extras (open range, unlock date, or maturity period).

5.7 Constant (Non-Varying) Provenance Fields

For completeness, the following fields are identical across all 2,448 rows — they describe the dataset's single generation pipeline rather than per-record variation:

Field	Constant value
source	ShareHub
source_name	sharehub_faq
source_repo	ShareHub
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	"" (empty)
metadata_json.generation_domain	वित्त (Finance)
metadata_json.generation_category	शेयर बजार सम्बन्धी प्रायः सोधिने प्रश्न (FAQ about the stock market)
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 (Nepali)
metadata_json.content_script	देवनागरी (Devanagari)
metadata_json.english_content_allowed	false
conversations length	Always 2 turns (human, gpt)

---

6. Text Length Statistics

Metric	Question (human turn)	Answer (gpt turn)
Minimum length (characters)	43	122
Maximum length (characters)	122	321
Mean length (characters)	81.0	236.7

Answers are, on average, roughly 3x longer than questions — consistent with the "informative answer" behavior, where the model expands a short factual question into a data-rich Nepali paragraph (often including exact figures, units such as अरब/लाख "billion/lakh," and derived ratios like price-to-book).

---

7. Naming Convention of id / source_row_id

Component	Pattern	Example
Prefix	sharehub_	sharehub_
Security slug	Lowercased ticker symbol	aclbsl, adblb, ymhl
Index suffix	Zero-padded 3-digit sequence number (per security, per template)	001, 002, ... 007
Full example	—	sharehub_aclbsl_001, sharehub_ymhl_006

id and source_row_id are always identical strings in this dataset.

---

8. License & Usage

Property	Value
License	Apache License 2.0
License tier	Permissive
Commercial use	Permitted under Apache-2.0 terms
Attribution	Recommended: cite "ShareHub" as the original data source

Refer to the full Apache-2.0 license text for exact terms and conditions before redistribution or commercial use.

---

9. How to Load This Dataset

The file is standard JSON Lines (.jsonl) — one self-contained JSON object per line, UTF-8 encoded (Devanagari text). Below are several equivalent ways to load it in Python.

9.1 Plain Python (no dependencies)

import json

path = "sharehub_faq_nepali__5_.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")

# Inspect the first record
first = records[0]
print("ID:", first["id"])
print("Question:", first["conversations"][0]["value"])
print("Answer:", first["conversations"][1]["value"])

# Parse the nested metadata_json string into a dict
metadata = json.loads(first["metadata_json"])
print("Security type:", metadata["security_type"])
print("Sub-domain:", metadata["generation_sub_domain"])
print("Original symbol:", metadata["original_symbol"])

9.2 Using pandas

import json
import pandas as pd

path = "sharehub_faq_nepali__5_.jsonl"

# Read JSONL directly into a DataFrame
df = pd.read_json(path, 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)

# Extract question / answer text into their own columns
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", "original_symbol", "security_type", "generation_sub_domain"]].head())

# Example: filter to only debenture questions about maturity period
maturity_qs = df[
    (df["security_type"] == "debenture") &
    (df["question"].str.contains("परिपक्वता"))
]
print(f"Maturity-period debenture questions: {len(maturity_qs)}")

# Example: value counts by sub-domain
print(df["generation_sub_domain"].value_counts())

9.3 Using jsonlines library

# pip install jsonlines
import jsonlines

path = "sharehub_faq_nepali__5_.jsonl"

records = []
with jsonlines.open(path) as reader:
    for obj in reader:
        records.append(obj)

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

9.4 Using Hugging Face datasets

# pip install datasets
from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files="sharehub_faq_nepali__5_.jsonl",
    split="train"
)

print(ds)
print(ds[0])

# Filter examples by security type using metadata_json (string field requires parsing)
import json

def get_security_type(example):
    return json.loads(example["metadata_json"])["security_type"]

ds = ds.map(lambda ex: {"security_type": get_security_type(ex)})
stocks_only = ds.filter(lambda ex: ex["security_type"] == "stock")
print(f"Stock-only rows: {len(stocks_only)}")

9.5 Streaming Line-by-Line (for very large files / low memory)

import json

path = "sharehub_faq_nepali__5_.jsonl"

def iter_records(path):
    with open(path, "r", encoding="utf-8") as f:
        for line in f:
            line = line.strip()
            if line:
                yield json.loads(line)

for i, record in enumerate(iter_records(path)):
    q = record["conversations"][0]["value"]
    a = record["conversations"][1]["value"]
    # process record here
    if i < 3:
        print(f"[{record['id']}] Q: {q}\n         A: {a}\n")

9.6 Quick Reproduction of the Statistics in This README

import json, collections, statistics

path = "sharehub_faq_nepali__5_.jsonl"
rows = [json.loads(l) for l in open(path, encoding="utf-8") if l.strip()]
metas = [json.loads(r["metadata_json"]) for r in rows]

print("Total rows:", len(rows))
print("Security type counts:", collections.Counter(m["security_type"] for m in metas))
print("Sub-domain counts:", collections.Counter(m["generation_sub_domain"] for m in metas))
print("Unique securities:", len(set(m["original_symbol"] for m in metas)))

qlens = [len(r["conversations"][0]["value"]) for r in rows]
alens = [len(r["conversations"][1]["value"]) for r in rows]
print("Mean question length:", round(statistics.mean(qlens), 1))
print("Mean answer length:", round(statistics.mean(alens), 1))

---

10. Suggested Use Cases

Use case	Why this dataset fits
Fine-tuning a Nepali financial FAQ assistant	Real, fact-grounded Q&A pairs about NEPSE-listed securities in natural Nepali
Evaluating LLM factual accuracy in low-resource languages	Precise numeric facts (market cap, book value, price-to-book, face value) in Devanagari script that can be checked against ground truth
Building a Nepali financial-domain retrieval/RAG corpus	Structured per-security facts across 8 consistent question templates
Studying instruction-tuning data for regional languages	Clean single-turn instruction/response pairs with rich metadata for filtering by domain, security type, or template
Domain classification / metadata tagging experiments	15 sub-domains and 3 security types provide ready-made labels

---

11. Caveats & Notes

• All rows share the same behavior, question_type, generation_domain, and generation_category labels — these fields do not vary and should not be treated as classification targets; the real signal for diversity is in generation_sub_domain, security_type, original_symbol, and the underlying question template (see §5.3).
• The url field is empty for every row in this file — no direct source URLs are included.
• Numbers inside answers are written using the Nepali numbering convention (लाख "lakh" = 100,000; अरब "arab" = 1,000,000,000) alongside Devanagari digits.
• english_content_allowed is false for all rows — company/instrument names may still appear in Latin script (as they are proper nouns/tickers), but no English explanatory text is present.
• Some securities have fewer than 5 associated questions (11 securities have only 1 question) — likely due to missing data fields (e.g., no recorded open range or promoter unlock date) at generation time.