sabin1234/NRB_Inflation_Annual_Report_FAQ_Dataset

NRB Inflation & Annual Report FAQ Dataset (Nepali)

A Nepali-language (Devanagari script) instruction-following dataset built from Nepal Rastra Bank (NRB) official statistical publications — specifically Annual Report economic-indicator datasets and the Inflation Indicators report for Fiscal Year 2079/80 (2022/23). Each record is a single-turn human → gpt FAQ pair where the question asks for a specific statistic (e.g., a growth rate, an index value, an inflation rate, an export/import/reserve figure for a given fiscal year), and the answer states that exact figure.

• File name: merged_nrb_inflation_full_faq_ne.jsonl
• Format: JSON Lines (.jsonl) — one JSON object per line
• Total records: 771
• Language: Nepali (ne / ISO 639-3 npi), Devanagari script
• License: CC0: Public Domain
• Task type: Instruction-following (single-turn Q&A)
• Generation type: Real (extracted from genuine NRB statistical publications — not synthetically generated)

---

---

1. Dataset Summary

Property	Value
Total records (rows)	771
Total unique sources (publications)	2
Total unique source_repo (publisher)	1 (Nepal Rastra Bank)
Total unique FAQ categories (generation_category)	2
Total unique domains (generation_domain)	2
Total unique sub-domains (generation_sub_domain)	3
Conversation turns per record	2 (1 human + 1 gpt), 100% of records
Language	Nepali (ne / npi)
Script	Devanagari (Deva)
License	CC0: Public Domain
License tier	Permissive
Task type	instruction-following
Generation type	real (100%)
Condition	real (100%)
English content allowed	False (100%)
Duplicate IDs	0 (all 771 IDs are unique)
Empty/missing url field	771 (100% — no source URL recorded)
source_config	default (100%)
source_split	train (100%)
source_revision	v1 (100%)
Distinct question templates (after removing fiscal-year prefix)	257
Distinct fiscal years referenced	15 (from 2008/09 to 2023/24)

---

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	"nrb_infl_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 specific publication the FAQ was derived from	"Inflation Indicators in Nepal for Fiscal Year 2079/80 (2022/23)"
source_name	string	Machine-readable slug identifying the specific publication/dataset	"inflation_indicators_nepal"
source_repo	string	Publishing organization (constant: Nepal Rastra Bank for all records)	"Nepal Rastra Bank"
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)	"nrb_infl_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	"CC0: Public Domain"
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": "खाद्य तथा पेय को वार्षिक मुद्रास्फीति दर विघटन कति थियो?" },
  { "from": "gpt",   "value": "खाद्य तथा पेय को वार्षिक मुद्रास्फीति दर विघटन २.८१ थियो।" }
]

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 (a statistic-lookup question) followed by one gpt turn (the exact numeric/factual answer). 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	"मुद्रास्फीति" (Inflation) or "मौद्रिक तथा वित्तीय" (Monetary & Financial)
generation_category	string	The specific publication/report this record belongs to	"मुद्रास्फीति सूचक" (Inflation Indicators)
generation_sub_domain	string	A finer-grained topical grouping	"नेपाल राष्ट्र बैंक" (Nepal Rastra Bank) or "आर्थिक सूचक" (Economic Indicators)
behavior	string	The intended response behavior/style label for the answer	"तथ्यमा आधारित उत्तर दिने" (Give a fact-based answer)
behavior_definition	string	A one-line definition explaining what the behavior label means	"मूल तथ्य कायम राख्दै शुद्ध नेपालीमा 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 statistical FAQ content from 2 distinct NRB publications, both published by the same institution (Nepal Rastra Bank):

Source (publication)	source_name (slug)	Record count	% of dataset	Content type
Nepal Rastra Bank Annual Report Datasets	nrb_annual_report	632	82.0%	Multi-year economic indicators (indices, growth rates, trade, reserves) extracted from NRB's Annual Report time-series
Inflation Indicators in Nepal for Fiscal Year 2079/80 (2022/23)	inflation_indicators_nepal	139	18.0%	Single fiscal-year inflation breakdown (annual inflation rate and percentage contribution, by consumption category)
Total		771	100%

Note: source_repo is constant ("Nepal Rastra Bank") for all 771 records — this is a single-publisher dataset, unlike multi-institution FAQ collections.

---

5. Domain & Sub-domain Diversity

Domains (generation_domain)

Domain (Nepali)	English meaning	Record count	% of dataset
मौद्रिक तथा वित्तीय	Monetary & Financial	632	82.0%
मुद्रास्फीति	Inflation	139	18.0%

Categories (generation_category)

Category (Nepali)	English meaning	Record count
नेपाल राष्ट्र बैंक वार्षिक प्रतिवेदन	Nepal Rastra Bank Annual Report	632
मुद्रास्फीति सूचक	Inflation Indicators	139

Sub-domains (generation_sub_domain)

Sub-domain (Nepali)	English meaning	Record count	% of dataset
आर्थिक सूचक	Economic Indicators	526	68.2%
विदेशी विनिमय रिजर्भ तथा आर्थिक सूचक	Foreign Exchange Reserves & Economic Indicators	106	13.7%
नेपाल राष्ट्र बैंक	Nepal Rastra Bank (inflation-report grouping)	139	18.0%
Total		771	100%

Cross-tabulation: Source × Sub-domain

Source	Sub-domain	Record count
Nepal Rastra Bank Annual Report Datasets	आर्थिक सूचक (Economic Indicators)	526
Nepal Rastra Bank Annual Report Datasets	विदेशी विनिमय रिजर्भ तथा आर्थिक सूचक (FX Reserves & Economic Indicators)	106
Inflation Indicators in Nepal for Fiscal Year 2079/80 (2022/23)	नेपाल राष्ट्र बैंक (Nepal Rastra Bank)	139

---

6. Behavior Diversity

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

This dataset uses a single, uniform behavior label across all 771 records. It is a purely factual, statistic-retrieval task — the model is expected to state an exact figure from an official report, with no interpretation, opinion, hedging, or elaboration.

behavior (Nepali)	English meaning	behavior_definition (Nepali)	English meaning	Record count	% of dataset
तथ्यमा आधारित उत्तर दिने	Give a fact-based answer	मूल तथ्य कायम राख्दै शुद्ध नेपालीमा उत्तर दिने।	Provide the answer in pure/standard Nepali while strictly preserving the original underlying fact/figure.	771	100%

There is zero behavioral diversity in this file — every answer is a direct, unembellished statement of a number or figure. If your training/evaluation pipeline needs varied response behaviors (e.g. explanatory, comparative, cautionary, conversational), you will need to combine this dataset with other behavior-diverse sources.

---

7. Question Pattern Diversity

The question_type metadata is likewise constant (तथ्यमा आधारित प्रश्न — "fact-based question") for all 771 records. However, the surface phrasing of questions falls into a small number of highly templated patterns, since this dataset was generated from structured statistical tables rather than free-form FAQ pages.

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
... कति थियो?	"How much was X?" (quantity/statistic lookup, past tense)	749	97.1%
... थियो? (sentence-final "was")	Confirms the question asks about a historical/reported value	703	91.2%
कसरी	"How" (present in a small number of process-style questions)	10	1.3%

Question ending punctuation style

Ending style	Record count	% of dataset
Ends with ? (question mark)	759	98.4%
Ends with । (Devanagari full stop, statement-style entries)	12	1.6%

Question structural template

The overwhelming majority of questions follow one of these two templates:

Template	Example	Approx. share
[YYYY/YY] मा [metric] कति थियो? — "In fiscal year YYYY/YY, how much was [metric]?"	"२०१२/१३ मा समग्र सूचकांक अन्तर्गत समग्र को मूल्य सूचकांक कति थियो?"	~522 records (67.7%)
[metric] कति थियो? (no explicit year prefix, typically referring to the single most recent reporting period)	"खाद्य तथा पेय को वार्षिक मुद्रास्फीति दर विघटन कति थियो?"	~139 records (18.0%, all from the Inflation Indicators source)

Question intent categories (qualitative grouping)

Intent category	Description	Approx. share
Time-series statistic lookup	"In year X, what was the value of indicator Y?"	~68%
Current-period statistic lookup	"What is/was the current value of indicator Y?" (no year specified, refers to latest available period)	~18%
Reserve/currency-denominated lookup	Foreign exchange reserve figures, broken out by currency (NPR vs. USD) and by reserve category	~14%

Key observation: Unlike a conversational FAQ dataset, this dataset's question diversity comes almost entirely from varying the metric name and fiscal year being asked about, not from varying the grammatical question form. Nearly all questions reduce to the single template "[what/when] + [metric] + कति थियो?" ("how much was it?").

---

8. Fiscal Year Coverage

Questions that include an explicit fiscal-year prefix (YYYY/YY मा ...) span 15 distinct Nepali fiscal years, from 2008/09 (२००८/०९) to 2023/24 (२०२३/२४):

Fiscal Year (B.S.)	Record count
२००८/०९ (2008/09)	12
२०१०/११ (2010/11)	12
२०११/१२ (2011/12)	12
२०१२/१३ (2012/13)	47
२०१३/१४ (2013/14)	34
२०१४/१५ (2014/15)	56
२०१५/१६ (2015/16)	19
२०१६/१७ (2016/17)	19
२०१७/१८ (2017/18)	40
२०१८/१९ (2018/19)	46
२०१९/२० (2019/20)	46
२०२०/२१ (2020/21)	46
२०२१/२२ (2021/22)	48
२०२२/२३ (2022/23)	47
२०२३/२४ (2023/24)	38
Records with a year prefix	522 (67.7% of dataset)
Records without a year prefix (single-period inflation-indicator questions)	249 (32.3% of dataset)

Coverage is uneven across years — earlier years (2008/09–2011/12) have far fewer records (12 each) than more recent years (2012/13 onward, 19–56 each), reflecting how many indicator line-items NRB reported in each year's Annual Report.

---

9. Economic Indicator / Metric Diversity

After stripping the fiscal-year prefix, there are 257 distinct question templates, corresponding to 257 distinct economic metrics/line-items. The most frequently repeated metrics (i.e., asked about across the most fiscal years) are:

Metric (Nepali)	English meaning	Times asked (across years)
नेपाल राष्ट्र बैंक को विदेशी विनिमय रिजर्भ (नेपाली रुपैयाँ)	Nepal Rastra Bank's foreign exchange reserve (NPR)	17
कुल रिजर्भ को विदेशी विनिमय रिजर्भ (नेपाली रुपैयाँ)	Total reserve's foreign exchange reserve (NPR)	17
नेपाल राष्ट्र बैंक को विदेशी विनिमय रिजर्भ (अमेरिकी डलर)	Nepal Rastra Bank's foreign exchange reserve (USD)	17
कुल रिजर्भ को विदेशी विनिमय रिजर्भ (अमेरिकी डलर)	Total reserve's foreign exchange reserve (USD)	17
खानी तथा उत्खनन को वृद्धि दर	Mining & quarrying growth rate	11
निर्माण उद्योग को वृद्धि दर	Manufacturing industry growth rate	11
निर्माण को वृद्धि दर	Construction growth rate	11
थोक तथा खुद्रा व्यापार को वृद्धि दर	Wholesale & retail trade growth rate	11
सार्वजनिक प्रशासन तथा रक्षा को वृद्धि दर	Public administration & defense growth rate	11
शिक्षा को वृद्धि दर	Education growth rate	11
कुल निर्यात	Total exports	10
भारततर्फ निर्यात	Exports to India	10
अन्य मुलुकतर्फ निर्यात	Exports to other countries	10
कुल आयात	Total imports	10
भारतबाट आयात	Imports from India	10
अन्य मुलुकबाट आयात	Imports from other countries	10
कुल व्यापार सन्तुलन	Total trade balance	10
भारतसँग व्यापार सन्तुलन	Trade balance with India	10
कुल वैदेशिक व्यापार	Total foreign trade	10

Metric category groups (qualitative)

Category	Examples	Approx. count of distinct templates
Sectoral GDP growth rates (वृद्धि दर)	Agriculture, mining, manufacturing, construction, education, health, transport, ICT, real estate, public administration, etc.	~30 sectors
Foreign trade	Total/India/other-country exports, imports, trade balance, foreign trade	~9 metrics
Foreign exchange reserves	NRB reserve, total (commercial bank + NRB) reserve — each in NPR and USD	~4 metrics
Price/consumption indices	Overall index, food & beverage index, housing & utilities index, transport index, etc.	Remaining templates
Inflation-specific (from inflation_indicators_nepal source only)	Annual inflation rate contribution & percentage contribution, per consumption category (food, housing, transport, etc.)	~139 records, ~70 distinct categories × 2 metrics (rate + contribution)

---

10. Conversation Structure Statistics

Property	Value
Records with exactly 2 turns (1 human + 1 gpt)	771 (100%)
Records with role order human → gpt	771 (100%)
Records with any other turn count or role order	0
Duplicate id values	0
id == source_row_id for all records	Yes (100%)

---

11. Text Length Statistics

Character-length statistics computed over the raw value strings (Devanagari text).

Turn	Min length (chars)	Max length (chars)	Average length (chars)
human (question)	27	90	53.9
gpt (answer)	26	92	58.0

Unlike conversational FAQ datasets, questions and answers here are almost the same length (average 53.9 vs. 58.0 characters) — the answer is essentially the question restated as a declarative statement with the number filled in (e.g., "X कति थियो?" → "X [number] थियो।"). There is no long-form explanation in either turn.

---

12. Sample Records

Inflation Indicators source (inflation_indicators_nepal):

{
  "id": "nrb_infl_faq_000001",
  "conversations": [
    {"from": "human", "value": "खाद्य तथा पेय को वार्षिक मुद्रास्फीति दर विघटन कति थियो?"},
    {"from": "gpt", "value": "खाद्य तथा पेय को वार्षिक मुद्रास्फीति दर विघटन २.८१ थियो।"}
  ],
  "source": "Inflation Indicators in Nepal for Fiscal Year 2079/80 (2022/23)",
  "source_name": "inflation_indicators_nepal",
  "source_repo": "Nepal Rastra Bank",
  "license": "CC0: Public Domain"
}

Annual Report source (nrb_annual_report), with fiscal-year prefix:

{
  "conversations": [
    {"from": "human", "value": "२०१२/१३ मा समग्र सूचकांक अन्तर्गत समग्र को मूल्य सूचकांक कति थियो?"},
    {"from": "gpt", "value": "२०१२/१३ मा समग्र सूचकांक अन्तर्गत समग्र को मूल्य सूचकांक १८१.७ थियो।"}
  ],
  "source": "Nepal Rastra Bank Annual Report Datasets",
  "source_name": "nrb_annual_report",
  "source_repo": "Nepal Rastra Bank"
}

Annual Report source — growth-rate metric:

{
  "conversations": [
    {"from": "human", "value": "२०१३/१४ मा खाद्य तथा पेय अन्तर्गत समग्र को मूल्य सूचकांक कति थियो?"},
    {"from": "gpt", "value": "२०१३/१४ मा खाद्य तथा पेय अन्तर्गत समग्र को मूल्य सूचकांक २३८.८ थियो।"}
  ],
  "source": "Nepal Rastra Bank Annual Report Datasets"
}

---

13. How to Load the Dataset

13.1 Plain Python (standard library only)

import json

path = "merged_nrb_inflation_full_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 domain
first = records[0]
question = first["conversations"][0]["value"]
answer   = first["conversations"][1]["value"]
domain   = first["metadata"]["generation_domain"]

print("Q:", question)
print("A:", answer)
print("Domain:", domain)

13.2 Using pandas

import json
import pandas as pd

path = "merged_nrb_inflation_full_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 the Annual Report source only
ar_df = df[df["source_name"] == "nrb_annual_report"]
print(f"Annual Report records: {len(ar_df)}")

# Example: extract fiscal year from the question (where present)
df["fiscal_year"] = df["question"].str.extract(r"^(\S+/\S+) मा")
print(df["fiscal_year"].value_counts())

13.3 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_nrb_inflation_full_faq_ne.jsonl",
    split="train",
)

print(dataset)
print(dataset[0])

# Filter by source using .filter()
inflation_only = dataset.filter(lambda x: x["source_name"] == "inflation_indicators_nepal")
print(f"Inflation Indicator records: {len(inflation_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)

13.4 Using jsonlines package (streaming, memory-efficient)

import jsonlines

path = "merged_nrb_inflation_full_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])

---

14. Quick Data Exploration Recipes

import json, re
from collections import Counter

path = "merged_nrb_inflation_full_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
print(Counter(r["metadata"]["generation_sub_domain"] for r in records))

# 3) Extract fiscal year from each question, where present
def get_year(q):
    m = re.match(r"^(\S+/\S+) मा", q)
    return m.group(1) if m else None

years = Counter(get_year(r["conversations"][0]["value"]) for r in records)
print(years.most_common())

# 4) Find all distinct metrics (question templates with year removed)
def strip_year(q):
    q = re.sub(r"^\S+ मा ", "", q)
    return q.replace(" कति थियो?", "").strip()

metrics = Counter(strip_year(r["conversations"][0]["value"]) for r in records)
print(f"Found {len(metrics)} distinct metric templates")
print(metrics.most_common(10))

# 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])

# 6) Extract the numeric answer value from each gpt turn (useful for building a structured table)
def extract_number(a):
    m = re.search(r"([\d१२३४५६७८९०.,\-]+)\s*(?:थियो|रह्यो)", a)
    return m.group(1) if m else None

numeric_answers = [extract_number(r["conversations"][1]["value"]) for r in records]
print("Sample extracted numeric answers:", numeric_answers[:10])

---

15. Data Quality Notes / Known Limitations

Note	Detail
Single behavior class	Every record shares one behavior label (तथ्यमा आधारित उत्तर दिने) — no explanatory, comparative, or conversational behavior variety exists in this file.
Single question-type label	The metadata question_type field is constant across all records (तथ्यमा आधारित प्रश्न); nearly all questions also share the same surface grammatical template ("... कति थियो?").
Highly templated phrasing	97.1% of questions follow the "[metric] कति थियो?" pattern — question diversity in this dataset comes from varying the metric name and fiscal year, not from varied grammatical question forms (contrast with more conversational FAQ datasets).
No source URLs	The url field is empty ("") for all 771 records — original page/report links were not retained.
Single publisher	source_repo is constant (Nepal Rastra Bank) — this dataset does not span multiple institutions.
Uneven fiscal-year coverage	Some years (e.g., 2008/09–2011/12) have only 12 records each, while more recent years (2012/13 onward) have 19–56 records — reflects how many line items NRB published per year, not a deliberate sampling choice.
Short, numeric-heavy answers	Answers restate the question as a declarative sentence with a number filled in — there is no elaboration, context, or explanation of what the figure means or why it changed.
Nepali-only, no code-switching	english_content_allowed is False for all records — content is Devanagari-only (including Devanagari numerals in most figures), so this dataset is not suited for code-switched (Nepali-English) evaluation.
Single-turn only	Every record is exactly one Q→A pair — this dataset alone cannot be used for multi-turn dialogue or follow-up-question 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.

---

16. Suggested Use Cases

• Fine-tuning or evaluating Nepali-language LLMs on precise, numeric fact-retrieval from official statistical reports.
• Building a retrieval-augmented generation (RAG) system or lookup tool over NRB's historical economic time-series (GDP growth by sector, trade balances, FX reserves, inflation breakdown).
• Testing a model's ability to read Devanagari numerals accurately and reproduce exact figures without hallucination.
• As a "hard factual grounding" component in a broader Nepali instruction-tuning mixture that also includes more conversational/explanatory finance FAQ data (e.g. to balance templated statistic-lookup questions against free-form Q&A).
• Constructing a structured (tabular) time-series dataset of Nepal's economic indicators by parsing out the fiscal year and metric from each question, and the numeric value from each answer (see Section 14, recipe #6).

---

17. License & Attribution

• License: CC0: Public Domain — see the license and license_tier fields in every record.
• Content origin: Real, publicly published statistical data from Nepal Rastra Bank's Annual Report series and its Inflation Indicators report for Fiscal Year 2079/80 (2022/23).
• As CC0 content, this data may be used, modified, and redistributed without restriction; however, consider noting Nepal Rastra Bank as the original data source for transparency and reproducibility.