sabin1234/Trade_Balance_Forest_Products_Comparative_FAQ_Dataset

Trade Balance & Forest Products Comparative FAQ Dataset (Nepali)

A Nepali-language (Devanagari script) instruction-following dataset built from two real Nepali government/official statistical publications: Nepal's Trade Balance by Partner Countries (FY 2074/75) and Forest Production Data Nepal (published by the Department of National Parks and Wildlife Conservation). Unlike simple single-value statistic lookups, almost every record in this dataset asks the model to compare two (or three) entities — two trading-partner countries, two fiscal years, or two forest products — and state which one is larger/smaller, or what the numeric difference between them is.

• File name: merged_trade_forest_faq_ne.jsonl
• Format: JSON Lines (.jsonl) — one JSON object per line
• Total records: 590
• Language: Nepali (ne / ISO 639-3 npi), Devanagari script
• License: CC-BY
• Task type: Instruction-following (single-turn, comparative Q&A)
• Generation type: Real (built from genuine trade and forestry statistics — not synthetically generated)

---

---

1. Dataset Summary

Property	Value
Total records (rows)	590
Total unique sources (publications)	2
Total unique source_repo (publisher/report title)	2
Total unique FAQ categories (generation_category)	2
Total unique domains (generation_domain)	2
Total unique sub-domains (generation_sub_domain)	36
Total unique behavior_definition variants	4
Total unique question_type variants	3
Conversation turns per record	2 (1 human + 1 gpt), 100% of records
Language	Nepali (ne / npi)
Script	Devanagari (Deva)
License	cc-by
License tier	Permissive
Task type	instruction-following
Generation type	real (100%)
Condition	real (100%)
English content allowed	False (100%)
Duplicate IDs	0 (all 590 IDs are unique)
Empty/missing url field	590 (100% — no source URL recorded)
source_config	default (100%)
source_split	train (100%)
source_revision	v1 (100%)
Records ending in ?	590 (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	"trade_forest_faq_000001"
conversations	array of objects	The single-turn dialogue. Each item has from ("human" or "gpt") and value (text)	see below
source	string	Broad source label ("other" for trade records, or the specific government body for forestry records)	"other" / "Department of National Parks and Wildlife Conservation"
source_name	string	Machine-readable slug identifying the specific dataset	"trade_balance_partner" / "forest_products_supply"
source_repo	string	The specific report/publication title the figures were sourced from	"Trade Balance by Partner Countries FY 2074/75" / "Forest Production Data Nepal"
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)	"trade_forest_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	"cc-by"
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": "भारत र चीनबाट आयात कति-कति छ र कुनबाट बढी आयात भयो?" },
  { "from": "gpt",   "value": "भारतबाट आयात 49662178.55 हजार रुपैयाँ छ भने चीनबाट 10819572.31 हजार रुपैयाँ छ। भारतबाट बढी आयात भएको छ।" }
]

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 record has exactly 2 turns: one human turn (a comparative or single-value statistic question) followed by one gpt turn (the answer, which usually restates both values and explicitly names the larger/smaller one). There are no multi-turn conversations.

---

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	"अन्तर्राष्ट्रिय व्यापार" (International Trade) or "वन तथा वन्यजन्तु" (Forest & Wildlife)
generation_category	string	The specific publication/report this record belongs to	"व्यापार सन्तुलन तथ्याङ्क" (Trade Balance Statistics) or "वन पैदावार आपूर्ति तथ्याङ्क" (Forest Produce Supply Statistics)
generation_sub_domain	string	A fine-grained comparison-type label — 36 distinct values, see Section 5	"आयात तुलना" (Import Comparison)
behavior	string	The intended response behavior/style label for the answer	"जानकारीमूलक उत्तर दिने" (Give an informative answer)
behavior_definition	string	A more detailed definition of the expected answer behavior — varies by comparison type (see Section 6)	"डेटासेटमा भएको वास्तविक अंकअनुसार शुद्ध नेपालीमा उत्तर दिने।"
question_type	string	The category of question being asked — 3 distinct values, see Section 7	"तुलनात्मक वा तथ्यमा आधारित प्रश्न" (Comparative or 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 (numeric figures are written in Latin/Arabic digits regardless)	false

---

4. Source Diversity

The dataset merges comparative-statistic FAQ content from 2 distinct official publications:

Source (publication)	source_name (slug)	source_repo	Record count	% of dataset	Content type
Trade Balance by Partner Countries FY 2074/75	trade_balance_partner	"Trade Balance by Partner Countries FY 2074/75"	570	96.6%	Bilateral trade figures (imports, exports, trade balance) between Nepal and its trading-partner countries
Forest Production Data Nepal	forest_products_supply	"Forest Production Data Nepal"	20	3.4%	Annual production volumes of forest products (timber, firewood, herbs, resin, etc.) by fiscal year
Total			590	100%

Note: the top-level source field is set to the generic label "other" for all 570 trade-balance records, while forestry records carry the specific government body name ("Department of National Parks and Wildlife Conservation"). Use source_name or source_repo rather than source if you need to reliably distinguish the two publications.

---

5. Domain & Sub-domain Diversity

Domains (generation_domain)

Domain (Nepali)	English meaning	Record count	% of dataset
अन्तर्राष्ट्रिय व्यापार	International Trade	570	96.6%
वन तथा वन्यजन्तु	Forest & Wildlife	20	3.4%

Categories (generation_category)

Category (Nepali)	English meaning	Record count
व्यापार सन्तुलन तथ्याङ्क	Trade Balance Statistics	570
वन पैदावार आपूर्ति तथ्याङ्क	Forest Produce Supply Statistics	20

Sub-domains (generation_sub_domain) — full breakdown, all 36 values

This is the field with the richest diversity in the dataset — it encodes the specific type of comparison being asked. Trade-balance sub-domains compare countries; forest-production sub-domains compare fiscal years and/or product types.

Trade balance sub-domains (16 types, 570 records):

Sub-domain (Nepali)	English meaning	Record count	% of dataset
आयात फरक	Import difference (numeric gap between two countries)	90	15.3%
आयात तुलना	Import comparison (which of two countries Nepal imports more from)	47	8.0%
एकल देश आयात निर्यात	Single country's import & export together	47	8.0%
दूर देश आयात तुलना	Import comparison, geographically distant country pairs	47	8.0%
निर्यात तुलना	Export comparison (which of two countries Nepal exports more to)	46	7.8%
निर्यात फरक	Export difference (numeric gap)	46	7.8%
सन्तुलन जोडी	Trade-balance pair (both countries' balance side by side)	42	7.1%
घाटा बचत तुलना	Deficit vs. surplus comparison	31	5.3%
एकल आयात	Single-country import value only	31	5.3%
एकल निर्यात	Single-country export value only	31	5.3%
स्थिति	Overall trade "status" with one country (surplus/deficit statement)	24	4.1%
उच्च निम्न घाटा तुलना	High vs. low deficit comparison	20	3.4%
बचत वा घाटा	Surplus-or-deficit yes/no framing	19	3.2%
तीन देश आयात	Three-way import comparison	18	3.1%
बचत देशहरू	Countries with a trade surplus	17	2.9%
शून्य निर्यात	Zero-export case (countries with no recorded exports)	14	2.4%

Forest production sub-domains (20 types, 20 records — one unique template per record):

Sub-domain (Nepali)	English meaning	Record count
हर्बल उत्पादन तुलना	Herbal product output comparison (2 years)	1
टिम्बर उत्पादन तुलना	Timber output comparison (2 years)	1
दाउरा उत्पादन तुलना	Firewood output comparison (2 years)	1
लौठ सल्ला उत्पादन तुलना	Resin (Lauth Salla) output comparison (2 years)	1
खयर उत्पादन तुलना	Khayar (Acacia catechu) output comparison (2 years)	1
अर्गेली उत्पादन तुलना	Argeli (Nepali paper plant) output comparison (2 years)	1
खोटो उत्पादन तुलना	Khoto output comparison (2 years)	1
लोक्ता उत्पादन तुलना	Lokta (Nepali paper plant) output comparison (2 years)	1
अन्य वन पैदावार तुलना	Other forest produce output comparison (2 years)	1
हर्बल र लोक्ता तुलना	Herbal vs. Lokta output comparison (same year)	1
टिम्बर र खयर तुलना	Timber vs. Khayar output comparison (same year)	1
खोटो र अर्गेली तुलना	Khoto vs. Argeli output comparison (same year)	1
हर्बल र खयर २००८/०९ तुलना	Herbal vs. Khayar, fiscal year 2008/09	1
टिम्बर र दाउरा २०११/१२	Timber vs. Firewood, fiscal year 2011/12	1
खोटो उत्पादन दुई वर्ष तुलना	Khoto output, two-year comparison	1
लौठ सल्ला दुई वर्ष तुलना	Resin output, two-year comparison	1
अर्गेली दुई वर्ष तुलना	Argeli output, two-year comparison	1
लोक्ता र अन्य वन पैदावार तुलना	Lokta vs. other forest produce comparison	1
हर्बल र खोटो २००५/०६	Herbal vs. Khoto, fiscal year 2005/06	1
टिम्बर दुई वर्ष तुलना	Timber output, two-year comparison	1

All 20 forest-production sub-domain labels are unique — each of the 20 forest records was tagged with its own descriptive label rather than being grouped into a shared category, giving this small slice of the dataset unusually fine-grained (essentially one-off) sub-domain tagging.

---

6. Behavior Diversity

Unlike the previous single-behavior-label datasets in this series, this dataset shows some behavior diversity through the behavior_definition field, even though the top-level behavior label itself is constant.

behavior (Nepali)	English meaning	Record count	% of dataset
जानकारीमूलक उत्तर दिने	Give an informative answer	590	100%

behavior_definition breakdown (4 distinct variants):

behavior_definition (Nepali)	English meaning	Record count	% of dataset	Applies to
डेटासेटमा भएको वास्तविक अंकअनुसार शुद्ध नेपालीमा उत्तर दिने।	Answer in pure/standard Nepali strictly using the real figures present in the dataset.	570	96.6%	All trade-balance records
डेटासेटमा भएको वास्तविक अंकअनुसार दुई वर्षको तुलना गरी शुद्ध नेपालीमा उत्तर दिने।	Answer in pure Nepali by comparing two fiscal years, using the real figures in the dataset.	13	2.2%	Forest records comparing the same product across two years
डेटासेटमा भएको वास्तविक अंकअनुसार दुई पैदावारको तुलना गरी शुद्ध नेपालीमा उत्तर दिने।	Answer in pure Nepali by comparing two different forest products, using the real figures in the dataset.	5	0.8%	Forest records comparing two different products in the same year
डेटासेटमा भएको वास्तविक अंकअनुसार दुई पैदावारको तथ्याङ्क शुद्ध नेपालीमा दिने।	State the statistics of two forest products in pure Nepali (no explicit comparison/verdict required).	2	0.3%	Forest records that only report both figures side-by-side without naming a "winner"

Key insight: while the trade-balance portion of the dataset (96.6%) uses one uniform behavior definition, the smaller forest-production portion (3.4%) is split three ways depending on what is being compared (two years of the same product vs. two different products vs. plain side-by-side reporting). This is the dataset's main axis of behavioral diversity.

---

7. Question Pattern Diversity

question_type metadata breakdown (3 distinct values)

question_type (Nepali)	English meaning	Record count	% of dataset
तुलनात्मक वा तथ्यमा आधारित प्रश्न	Comparative or fact-based question	570	96.6%
तुलनात्मक तथ्यमा आधारित प्रश्न	Comparative fact-based question	18	3.1%
तथ्यमा आधारित प्रश्न	Fact-based question (non-comparative)	2	0.3%

Surface question-marker breakdown

(A single question can match more than one marker, so percentages do not sum to 100%.)

Question pattern (Nepali marker)	English meaning	Matching records	% of dataset	Example
... कति छ?	"How much is X?" (present-tense quantity question)	349	59.2%	"भारत र चीनबाट आयात कति-कति छ...?"
कति-कति ... ?	Paired quantity question ("how much [is] each")	200	33.9%	"भारत र चीनबाट आयात कति-कति छ...?"
फरक कति ... ?	"What is the difference...?"	136	23.1%	"भारत र संयुक्त अरब इमिरेट्सबाट आयातको फरक कति छ?"
कुनबाट बढी ... ?	"From which one was it more?"	94	15.9%	"भारत र चीनबाट आयात कति-कति छ र कुनबाट बढी आयात भयो?"
बचत / घाटा	"Surplus" / "deficit" framing	87	14.7%	"भारत र चीनसँगको व्यापार घाटा/बचत कुन बढी छ?"
तुलना गर्दा ... ?	"When compared, ...?"	71	12.0%	"हर्बल र लोक्ताको २०१३/१४ को उत्पादन तुलना गर्दा कुन बढी थियो?"
कुन बढी / कुन धेरै ... ?	"Which one is/was more?"	56	9.5%	"टिम्बर र खयरको ... तुलना गर्दा कुन बढी थियो?"
... कति थियो?	"How much was X?" (past-tense quantity question)	16	2.7%	"टिम्बर उत्पादन २०१०/११ र २०१३/१४ मा कति थियो...?"
कुन वर्ष बढी/कम ... ?	"In which year was it more/less?"	12	2.0%	"टिम्बर उत्पादन ... कुन वर्ष कम थियो?"

Question structural templates

Template	Example	Domain
[Country A] र [Country B]बाट आयात कति-कति छ र कुनबाट बढी आयात भयो? — "How much does Nepal import from A and B, and from which is it higher?"	"भारत र चीनबाट आयात कति-कति छ र कुनबाट बढी आयात भयो?"	Import comparison
[Country A] र [Country B]सँग निर्यात कति-कति छ र कुनसँग बढी निर्यात भयो? — export equivalent	"चीन र फ्रान्ससँग निर्यात कति-कति छ र कुनसँग बढी निर्यात भयो?"	Export comparison
[Country A] र [Country B]सँगको व्यापार घाटा/बचत कुन बढी छ? — "Whose trade deficit/surplus with A vs B is larger?"	"भारत र चीनसँगको व्यापार घाटा/बचत कुन बढी छ?"	Deficit/surplus comparison
[Country]बाट आयात र निर्यात कति-कति छ? — "What are country X's import and export figures?"	"भारतबाट आयात र निर्यात कति-कति छ?"	Single-country import+export
[Country A] र [Country B]सँग निर्यात कति छ? (zero-export framing)	"अर्जेन्टिना र इजिप्टसँग निर्यात कति छ?"	Zero-export cases
[Country A], [Country B] र [Country C] मध्ये कुनबाट सबैभन्दा बढी आयात भएको छ? — three-way "which is highest"	"भारत, फ्रान्स र संयुक्त अरब इमिरेट्स मध्ये कुनबाट सबैभन्दा बढी आयात भएको छ?"	Three-country import ranking
[Product] उत्पादन [Year A] र [Year B] मा कति थियो र कुन वर्ष बढी/कम थियो? — "How much was product X produced in year A vs year B, and which year was higher/lower?"	"टिम्बर उत्पादन २०१०/११ र २०१३/१४ मा कति थियो र कुन वर्ष कम थियो?"	Forest product, year-over-year
[Product A] र [Product B]को [Year] को उत्पादन तुलना गर्दा कुन बढी थियो? — "Comparing products A and B's output in year Y, which was more?"	"हर्बल र लोक्ताको २०१३/१४ को उत्पादन तुलना गर्दा कुन बढी थियो?"	Forest product-to-product, same year

Question intent categories (qualitative grouping)

Intent category	Description	Approx. share
Pairwise comparison ("which is more/less?")	Two entities (countries, years, or products) compared, with a clear winner named	~55%
Numeric difference ("what is the gap?")	Explicitly asks for the arithmetic difference between two figures	~23%
Single-entity lookup	Asks for one country's or one product's figure without comparison	~11%
Three-way ranking	Asks which of three countries has the highest value	~3%
Zero/absence framing	Asks about a pair where one or both values are zero	~2%
Status/classification	Asks whether a relationship is a "surplus" or "deficit" rather than for a number	~3%

Key observation: This dataset has the richest question-pattern diversity of the three datasets examined so far in this series — question variety comes not just from varying the named entities (countries, years, products) but from varying the comparison structure itself (pairwise vs. three-way, difference vs. verdict, surplus/deficit framing vs. plain magnitude, zero-value edge cases, etc.).

---

8. Sub-domain × Source Cross-tabulation

source_name	Sub-domains present	Total records
trade_balance_partner	आयात तुलना (47), निर्यात तुलना (46), घाटा बचत तुलना (31), एकल देश आयात निर्यात (47), बचत देशहरू (17), शून्य निर्यात (14), उच्च निम्न घाटा तुलना (20), एकल आयात (31), एकल निर्यात (31), स्थिति (24), दूर देश आयात तुलना (47), सन्तुलन जोडी (42), आयात फरक (90), निर्यात फरक (46), बचत वा घाटा (19), तीन देश आयात (18)	570
forest_products_supply	20 distinct one-off sub-domain labels (see Section 5)	20

---

9. Conversation Structure Statistics

Property	Value
Records with exactly 2 turns (1 human + 1 gpt)	590 (100%)
Records with role order human → gpt	590 (100%)
Records with any other turn count or role order	0
Duplicate id values	0
id == source_row_id for all records	Yes (100%)

---

10. Text Length Statistics

Character-length statistics computed over the raw value strings (Devanagari text; numeric figures are written with Latin/Arabic digits within the Devanagari sentences).

Turn	Min length (chars)	Max length (chars)	Average length (chars)
human (question)	29	81	49.1
gpt (answer)	38	153	96.1

Answers are, on average, roughly 2x longer than questions — consistent with comparative answers that must state two separate numeric values (and the unit) plus an explicit comparison verdict sentence, rather than a single fact.

---

11. Sample Records

Trade balance — import comparison:

{
  "id": "trade_forest_faq_000001",
  "conversations": [
    {"from": "human", "value": "भारत र चीनबाट आयात कति-कति छ र कुनबाट बढी आयात भयो?"},
    {"from": "gpt", "value": "भारतबाट आयात 49662178.55 हजार रुपैयाँ छ भने चीनबाट 10819572.31 हजार रुपैयाँ छ। भारतबाट बढी आयात भएको छ।"}
  ],
  "source": "other",
  "source_name": "trade_balance_partner",
  "source_repo": "Trade Balance by Partner Countries FY 2074/75",
  "license": "cc-by"
}

Trade balance — zero-export edge case:

{
  "conversations": [
    {"from": "human", "value": "अर्जेन्टिना र इजिप्टसँग निर्यात कति छ?"},
    {"from": "gpt", "value": "अर्जेन्टिना र इजिप्ट दुवैसँग निर्यात शून्य (०) छ।"}
  ],
  "source_name": "trade_balance_partner"
}

Trade balance — three-way ranking:

{
  "conversations": [
    {"from": "human", "value": "भारत, फ्रान्स र संयुक्त अरब इमिरेट्स मध्ये कुनबाट सबैभन्दा बढी आयात भएको छ?"},
    {"from": "gpt", "value": "भारतबाट 49662178.55, फ्रान्सबाट 1661903.56 र संयुक्त अरब इमिरेट्सबाट 1397507.91 हजार रुपैयाँ आयात भएको छ। सबैभन्दा बढी भारतबाट आयात भएको छ।"}
  ],
  "source_name": "trade_balance_partner"
}

Forest products — year-over-year comparison:

{
  "conversations": [
    {"from": "human", "value": "२००४/०५ र २०१४/१५ मा हर्बल उत्पादन कति-कति थियो र कुन वर्षको उत्पादन बढी थियो?"},
    {"from": "gpt", "value": "२००४/०५ मा हर्बल उत्पादन ४५७५५७९ किलोग्राम थियो भने २०१४/१५ मा ३०७९०२३ किलोग्राम थियो। २००४/०५ को उत्पादन बढी थियो।"}
  ],
  "source": "Department of National Parks and Wildlife Conservation",
  "source_name": "forest_products_supply",
  "source_repo": "Forest Production Data Nepal"
}

Forest products — cross-product comparison with unit mismatch:

{
  "conversations": [
    {"from": "human", "value": "२००८/०९ मा हर्बल र खयरको उत्पादन कति थियो र कुन बढी थियो?"},
    {"from": "gpt", "value": "२००८/०९ मा हर्बल उत्पादन ११६५३४५ किलोग्राम थियो भने खयर उत्पादन ३२०५७ घन फिट थियो। एकाइ फरक भए पनि हर्बलको अंक बढी देखिन्छ।"}
  ],
  "source_name": "forest_products_supply"
}

Note this last example: herbal output is measured in kilograms while Khayar output is measured in cubic feet — the answer explicitly flags that the units differ ("एकाइ फरक भए पनि" — "even though the units differ") before stating which raw number is larger. This is a useful edge case to be aware of if you use this data for arithmetic-grounded fine-tuning, since the "comparison" here is not a true apples-to-apples comparison.

---

12. How to Load the Dataset

12.1 Plain Python (standard library only)

import json

path = "merged_trade_forest_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 sub-domain
first = records[0]
question = first["conversations"][0]["value"]
answer   = first["conversations"][1]["value"]
sub_domain = first["metadata"]["generation_sub_domain"]

print("Q:", question)
print("A:", answer)
print("Sub-domain:", sub_domain)

12.2 Using pandas

import json
import pandas as pd

path = "merged_trade_forest_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_name", "generation_sub_domain", "question", "answer"]].head())

# Example: filter records from the trade-balance source only
trade_df = df[df["source_name"] == "trade_balance_partner"]
print(f"Trade balance records: {len(trade_df)}")

# Example: group by sub-domain and count
print(df.groupby("generation_sub_domain").size().sort_values(ascending=False))

# Example: isolate forest-product records
forest_df = df[df["source_name"] == "forest_products_supply"]
print(forest_df[["question", "answer"]])

12.3 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_trade_forest_faq_ne.jsonl",
    split="train",
)

print(dataset)
print(dataset[0])

# Filter by source using .filter()
forest_only = dataset.filter(lambda x: x["source_name"] == "forest_products_supply")
print(f"Forest product records: {len(forest_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)

12.4 Using jsonlines package (streaming, memory-efficient)

import jsonlines

path = "merged_trade_forest_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])

---

13. Quick Data Exploration Recipes

import json, re
from collections import Counter

path = "merged_trade_forest_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_name"] for r in records))

# 2) Count records per sub-domain (all 36 values)
print(Counter(r["metadata"]["generation_sub_domain"] for r in records).most_common())

# 3) Count records per behavior_definition (the dataset's main behavior-diversity axis)
print(Counter(r["metadata"]["behavior_definition"] for r in records))

# 4) Find all "difference" (फरक) questions and their answers
diff_questions = [
    r for r in records
    if "फरक कति" in r["conversations"][0]["value"]
]
print(f"Found {len(diff_questions)} 'difference' questions")
print(diff_questions[0]["conversations"])

# 5) Find zero-export edge cases
zero_export = [
    r for r in records
    if "शून्य" in r["conversations"][1]["value"]
]
print(f"Found {len(zero_export)} zero-value records")

# 6) 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])

# 7) Extract all numeric values mentioned in each answer (for structured/tabular reuse)
def extract_numbers(a):
    return re.findall(r"[\d]+\.?\d*", a)

for r in records[:3]:
    print(r["conversations"][1]["value"], "->", extract_numbers(r["conversations"][1]["value"]))

---

14. Data Quality Notes / Known Limitations

Note	Detail
Two very differently-sized sub-datasets	Trade-balance records (570, 96.6%) vastly outnumber forest-production records (20, 3.4%) — treat the forest-production slice as a small supplementary set, not a balanced counterpart.
Ambiguous source field for trade records	The top-level source field is the generic string "other" for all trade-balance records; use source_name (trade_balance_partner) or source_repo to reliably identify this sub-source.
Highly templated phrasing per sub-domain	Within each of the 16 trade-balance sub-domains, phrasing is near-identical except for the substituted country names and numbers — diversity comes from entity substitution, not varied grammar (similar to the NRB inflation dataset in this series).
One-off sub-domain granularity for forest data	Each of the 20 forest-production records has its own unique generation_sub_domain label (e.g., "हर्बल उत्पादन तुलना" appears exactly once) — these labels function more like per-record descriptive tags than reusable categories, so don't expect to find multiple records under most forest sub-domain labels.
Unit-mismatch edge case	At least one sample (see Section 11) compares two products measured in different units (kilograms vs. cubic feet) and explicitly notes the mismatch rather than silently treating the numbers as comparable — worth flagging if you build automated correctness checks against the raw numbers.
No source URLs	The url field is empty ("") for all 590 records.
Nepali-only, no code-switching	english_content_allowed is False for all records — content is Devanagari-only (though numeric figures use Latin/Arabic digits), so this dataset is not suited for code-switched (Nepali-English) evaluation.
Single-turn only	Every record is exactly one Q→A pair — no multi-turn or follow-up-question data.
Fixed split/config/revision metadata	source_config, source_split, and source_revision are constant (default, train, v1) — carry no discriminative information in this file.
Single fiscal year for trade data	All trade-balance figures are drawn from a single fiscal year (2074/75), so this portion of the dataset cannot support year-over-year trade trend questions — contrast with the forest-production portion, which does span multiple years.

---

15. Suggested Use Cases

• Fine-tuning or evaluating Nepali-language LLMs on comparative reasoning over real numeric data (which of two values is larger, and by how much).
• Testing whether a model correctly reads large Devanagari-embedded numeric figures and performs basic arithmetic (differences, rankings) without hallucination.
• Building a retrieval/lookup tool over Nepal's bilateral trade-balance data (by partner country) or historical forest-product output data (by year/product).
• Stress-testing a model's handling of edge cases: zero values, unit mismatches between compared quantities, and three-way (not just two-way) comparisons.
• As the "comparative reasoning" component in a broader Nepali instruction-tuning mixture, complementing more purely descriptive or single-value statistic datasets (e.g., the NRB Annual Report / Inflation Indicators dataset in this same series).

---

16. License & Attribution

• License: cc-by (Creative Commons Attribution) — see the license and license_tier fields in every record.
• Content origin: Real, publicly published statistics from Nepal's Trade Balance by Partner Countries report (FY 2074/75) and the Department of National Parks and Wildlife Conservation's Forest Production Data Nepal report.
• CC-BY requires attribution to the original source when redistributing — credit Nepal's trade-statistics publisher and the Department of National Parks and Wildlife Conservation, respectively, for the two underlying reports.