sabin1234/Nepal_Economic_Statistics_FAQ

🇳🇵 Nepal Economic Statistics FAQ (Nepali) — Merged Dataset README

A merged, sequentially re-indexed dataset of 645 Nepali-language factual Q&A pairs covering three related economic domains: commodity/SITC trade, GDP by ISIC sector, and customs import/export by checkpoint. Built by combining three source datasets into a single JSONL file.

---

🔖 TL;DR

What	Value
Total records	645
Output file	merged_nepal_economic_stats.jsonl
File size	~1.21 MB
Format	JSONL, 1 JSON object per line
Conversation style	ShareGPT-style conversations list, 2 turns (human → gpt)
Language / Script	Nepali (ne / npi) — Devanagari
Task type	instruction-following
License	CC-BY (permissive), mixed casing cc-by / CC-BY
id format	nepal_economic_stats_001 → nepal_economic_stats_645 (zero-padded, sequential, no gaps)
Source files merged	3 (see table below)
Duplicate check	0 duplicate IDs, 0 duplicate questions, 0 duplicate answers

---

🗂️ How this file was built

Three separate Nepali-language JSONL datasets were concatenated in this exact order, and every record's id field was overwritten with a new sequential, zero-padded ID — everything else in each record (conversations, metadata, license, etc.) was left untouched.

Order	Source file	Records	New ID range
1	commodity_trade_sitc_nepali_clean.jsonl	275	nepal_economic_stats_001 → nepal_economic_stats_275
2	gdp_isic_2018_nepali_clean.jsonl	285	nepal_economic_stats_276 → nepal_economic_stats_560
3	nepal_customs_import_export_nepali_clean.jsonl	85	nepal_economic_stats_561 → nepal_economic_stats_645
	Total	645	nepal_economic_stats_001 → nepal_economic_stats_645

The original id values (e.g. commodity_trade_sitc_001, gdp_isic_001_01, nepal_customs_trade_english_001) used three different, inconsistent naming schemes. They have been replaced with one unified, serial scheme so the merged file can be indexed and referenced consistently. The original identifiers are still recoverable from the untouched source_row_id field inside each record.

---

🧬 Schema

Every line is an independent JSON object with the same top-level structure:

{
  "id": "nepal_economic_stats_001",
  "conversations": [
    {"from": "human", "value": "२००५/०६ मा कुल निर्यातको मूल्य कति हो?"},
    {"from": "gpt",   "value": "२००५/०६ मा कुल निर्यातको दर्ज मूल्य ६०,२३४.१ हो।"}
  ],
  "source": "other",
  "source_name": "commodity_trade_by_sitc_group_2005_06_to_2015_16",
  "source_repo": "other",
  "source_config": "default",
  "source_split": "train",
  "source_revision": "v1",
  "source_row_id": "commodity_trade_sitc_source_001",
  "language": "ne", "language_code": "npi", "script": "Deva",
  "license": "cc-by", "license_tier": "permissive",
  "task_type": "instruction-following",
  "generation_type": "real", "condition": "real",
  "url": "",
  "metadata_json": "{...}"
}

Field	Type	Notes
id	string	✅ New — unique, sequential, zero-padded (nepal_economic_stats_NNN)
conversations	list[2]	Always human → gpt, exactly 2 turns, no system prompt
source, source_name, source_repo	string	Varies by origin file — see below
source_row_id	string	Original pre-merge identifier, preserved for traceability
language / language_code / script	string	Constant: ne / npi / Deva across all 645
license / license_tier	string	Constant meaning (CC-BY, permissive), but casing differs by file (cc-by vs CC-BY)
task_type	string	Constant: instruction-following
generation_type / condition	string	Constant: real (grounded in real statistical tables, not synthetic)
metadata_json	string (nested JSON)	Domain/category/behavior tags — varies by source, see below
license_id, license_name	string	Only present on the 85 customs records (extra fields not found in the other two source files)

Note on schema consistency: the customs subset carries two extra fields (license_id, license_name) that the other two subsets don't have. This is a genuine difference inherited from the source files — no fields were added or removed during merging.

---

📦 The 3 sub-datasets inside this merge

Sub-dataset	Records	Domain (from metadata_json.generation_domain)	Category
Commodity Trade by SITC	275	अन्तर्राष्ट्रिय व्यापार (International Trade)	एसआईटीसी समूह अनुसार वस्तु व्यापार
GDP by ISIC Sector	285	आर्थिक तथ्याङ्क (Economic Statistics)	आईएसआईसी विभाजन अनुसार कुल गार्हस्थ उत्पादन संरचना
Customs Import/Export	85	वित्तीय सेवा (Financial Services)	भन्सार आयात तथा निर्यात तथ्याङ्क

Commodity Trade (SITC)  █████████████████████████████████████████ 42.6%  (275)
GDP by ISIC Sector      ████████████████████████████████████████████ 44.2%  (285)
Customs Import/Export   █████████████ 13.2%  (85)

---

1️⃣ Commodity Trade by SITC (275 records)

Questions about Nepal's import/export value by SITC commodity group, fiscal years 2005/06 – 2015/16 (11 fiscal years).

• 22 distinct commodity categories (11 export groups + 11 matching import groups), e.g. कुल निर्यात (Total Exports), खाद्य तथा जीवित जनावरको निर्यात (Food & Live Animals Exports), मेसिनरी तथा यातायात उपकरणको आयात (Machinery & Transport Equipment Imports), etc. Each category has ~12–13 records (one per fiscal year).
• behavior (from metadata) breaks into 3 sub-tasks:

Behavior	Count	%	Meaning
प्रत्यक्ष तथ्य पुनर्प्राप्ति (Direct fact retrieval)	234	85.1%	"What was the value in year X?"
उच्चतम मूल्य चयन (Highest-value selection)	22	8.0%	"Which year had the highest value?"
न्यूनतम मूल्य चयन (Lowest-value selection)	19	6.9%	"Which year had the lowest value?"

• This is the only sub-dataset with comparison/extremum-style questions (min/max across years) — not pure single-fact lookup like the other two subsets.
• Average length: Q ≈ 61 chars, A ≈ 66 chars — the shortest, most compact subset.

---

2️⃣ GDP by ISIC Sector (285 records)

Questions about each economic sector's share (%) of GDP at current prices, fiscal years 2000/01 – 2017/18 (18 fiscal years).

• 15 ISIC sectors covered: कृषि तथा वन (Agriculture & Forestry), मत्स्यपालन (Fisheries), खानी तथा उत्खनन (Mining), उत्पादन (Manufacturing), बिजुली/ग्यास/पानी (Electricity/Gas/Water), निर्माण (Construction), थोक तथा खुद्रा व्यापार (Wholesale & Retail Trade), होटल तथा रेस्टुराँ (Hotels & Restaurants), यातायात/भण्डारण/सञ्चार (Transport/Storage/Communication), वित्तीय मध्यस्थता (Financial Intermediation), जग्गा/भाडा/व्यावसायिक सेवा (Real Estate/Renting/Business), सार्वजनिक प्रशासन (Public Administration), शिक्षा (Education), स्वास्थ्य (Health), and अन्य सेवा (Other Services).
• Structure: 15 sectors × 18 years = 270 year-by-year records, plus 15 "18-year average" summary records (one per sector) = 285 total.
• 100% of records share one behavior: तथ्यात्मक उत्तर दिने (give a factual answer) — no min/max or comparison questions here, unlike the commodity subset.
• Average length: Q ≈ 87 chars, A ≈ 94 chars — the longest subset (sector names are long compound Nepali phrases).

---

3️⃣ Customs Import/Export (85 records)

Questions about import/export value by customs checkpoint, fiscal year 2074/75 (first month) only — a single-period snapshot across 15 checkpoints (14 named points + 1 national total).

Checkpoint	Records
वीरगञ्ज (Birgunj), भैरहवा (Bhairahawa), Dry Port, त्रि. वि. विमानस्थल (TIA Airport), विराटनगर (Biratnagar), नेपालगञ्ज (Nepalgunj), कृष्णनगर (Krishnanagar), मेची (Mechi), कैलाली (Kailali), रसुवा (Rasuwa), जलेश्वर (Jaleshwor), जनकपुर (Janakpur), कञ्चनपुर (Kanchanpur), सुनसरी (Sunsari), सिराहा (Siraha)	5 each (70 total)
कुल/जम्मा (National Total)	10

• Question-type breakdown:

Type	Count	%
Import value (रु.)	36	42.4%
Export value (रु.)	18	21.2%
Import as % of national total	16	18.8%
Export as % of national total	15	17.6%

• Average length: Q ≈ 65 chars, A ≈ 58 chars.
• This subset carries 2 extra schema fields (license_id, license_name) not present in the other two.

---

💬 Conversation & length statistics (whole merged file)

	Value
Turn count	Always exactly 2 (human → gpt), 645/645
Question length (chars)	min 31, max 147, avg 73.2
Answer length (chars)	min 39, max 149, avg 77.2
Unique id	645 / 645 ✅
Unique questions	645 / 645 ✅ (no duplicates)
Unique answers	645 / 645 ✅ (no duplicates)
Missing/empty fields	None found

Q length by subset:  Commodity ~61  |  GDP ~87  |  Customs ~65   (chars, avg)
A length by subset:  Commodity ~66  |  GDP ~94  |  Customs ~58   (chars, avg)

---

🧾 Field homogeneity (constant across all 645 records)

These fields never change, regardless of sub-dataset:

• language = ne, language_code = npi, script = Deva
• license_tier = permissive
• task_type = instruction-following
• generation_type = condition = real (grounded in real government/statistical tables, not synthetic/hallucinated)
• url = "" (always empty)
• question_type (inside metadata_json) = तथ्यमा आधारित प्रश्न (fact-based question) for all 645

These fields vary by sub-dataset of origin (not randomly, but consistently within each source):

• source, source_name, source_repo, source_row_id
• license casing (cc-by for commodity/GDP, CC-BY for customs)
• generation_domain, generation_category, generation_sub_domain, behavior (inside metadata_json)
• presence of license_id / license_name (customs only)

---

⚠️ Things to know before using this dataset

• Uneven length distribution across subsets — GDP questions/answers run noticeably longer (long compound sector names) than commodity or customs ones. If you're training or benchmarking, be aware of this length skew.
• Different time coverage per subset — commodity trade covers 2005/06–2015/16, GDP covers 2000/01–2017/18, customs covers only one fiscal-year snapshot (2074/75, first month). They are not directly time-aligned with each other.
• The commodity subset is the only one with comparison-style (min/max) questions — if you need pure single-fact lookups only, you may want to filter on metadata_json.behavior == "प्रत्यक्ष तथ्य पुनर्प्राप्ति".
• License casing is inconsistent (cc-by vs CC-BY) — normalize this field if you're filtering/grouping by license programmatically.
• id was rewritten; if you need to trace a record back to its original source file/ID, use source_row_id (unchanged) together with the ID-range table above.

---

💻 How to load this dataset

Option 1 — Plain Python (no dependencies)

import json

records = []
with open("merged_nepal_economic_stats.jsonl", 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]["id"])
print(records[0]["conversations"][0]["value"])  # question
print(records[0]["conversations"][1]["value"])  # answer

Option 2 — pandas

import pandas as pd

df = pd.read_json("merged_nepal_economic_stats.jsonl", lines=True)
print(df.shape)          # (645, N)
print(df["id"].head())

# Pull out 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[["id", "question", "answer"]].head())

Option 3 — 🤗 Hugging Face datasets

from datasets import load_dataset

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

# Flatten conversations into question/answer columns
def flatten(example):
    example["question"] = example["conversations"][0]["value"]
    example["answer"] = example["conversations"][1]["value"]
    return example

ds = ds.map(flatten)
print(ds[0]["question"], "->", ds[0]["answer"])

Option 4 — Filter by sub-dataset / behavior (using pandas)

import pandas as pd, json

df = pd.read_json("merged_nepal_economic_stats.jsonl", lines=True)

# Split back into the 3 original sub-datasets using the known ID ranges
def get_subset(id_str):
    n = int(id_str.split("_")[-1])
    if n <= 275: return "commodity_trade_sitc"
    if n <= 560: return "gdp_isic_2018"
    return "nepal_customs_import_export"

df["subset"] = df["id"].apply(get_subset)
print(df["subset"].value_counts())

# Extract behavior from metadata_json for further filtering
df["behavior"] = df["metadata_json"].apply(lambda m: json.loads(m).get("behavior"))
direct_facts = df[df["behavior"] == "प्रत्यक्ष तथ्य पुनर्प्राप्ति"]
print(f"Direct-fact-only records: {len(direct_facts)}")

---

🧪 Sample records (one from each sub-dataset)

{"id": "nepal_economic_stats_001",
 "conversations": [
   {"from": "human", "value": "२००५/०६ मा कुल निर्यातको मूल्य कति हो?"},
   {"from": "gpt", "value": "२००५/०६ मा कुल निर्यातको दर्ज मूल्य ६०,२३४.१ हो।"}
 ]}

{"id": "nepal_economic_stats_276",
 "conversations": [
   {"from": "human", "value": "२०००/०१ मा चालु मूल्यमा कुल गार्हस्थ उत्पादनमा कृषि तथा वनको अंश कति प्रतिशत थियो?"},
   {"from": "gpt", "value": "२०००/०१ मा चालु मूल्यमा कृषि तथा वनले कुल गार्हस्थ उत्पादनको ३६.१५% अंश ओगटेको थियो।"}
 ]}

{"id": "nepal_economic_stats_561",
 "conversations": [
   {"from": "human", "value": "आर्थिक वर्ष २०७४/७५ को पहिलो महिनामा वीरगञ्जमा दर्ज आयात मूल्य कति थियो?"},
   {"from": "gpt", "value": "वीरगञ्जमा आयात मूल्य रु. २६,१९५,०७१.९३ हजार दर्ज भएको थियो।"}
 ]}

---

📜 License

CC-BY (permissive, attribution required) across all 645 records — inherited unchanged from the three source files. Verify the specific license/license_id/license_name fields on the records you use if attribution requirements matter for your use case.

---

This README was generated from a full programmatic scan of the merged 645-record file — every number above is computed directly from the data, not estimated.