sabin1234/ROMANIZED_SCRAP_Finance_Instruction_Dataset

---

license: apache-2.0 language: - ne tags: - finance - nepali - romanized - banking - nepse - instruction-following - sharegpt - nepal-rastra-bank pretty_name: Romanized Nepali Finance Dataset size_categories: - 10K<n<100K

Romanized Nepali Finance Dataset

A large-scale Romanized Nepali instruction-following dataset covering Nepal's financial ecosystem — commercial banking indicators, NEPSE-listed company data, Nepal Rastra Bank BFI registry, and savings account statistics. All questions and answers are written in Romanized Nepali (Nepali language rendered in Latin script), making the dataset uniquely suited for training and evaluating language models on low-resource Nepali financial text.

---

Dataset Summary

Field	Value
Total Rows	23,855
Language	Nepali (ne)
Script	Romanized Latin (Latn) / Mixed
Language Code	nep
Format	ShareGPT (multi-turn conversation)
Conversation Turns	2 (1 human + 1 assistant) per sample
License	Apache-2.0
Generation Type	Real
Condition	Real
Domain	Finance, Banking, Capital Markets

---

Data Sources

The dataset is compiled from 4 distinct source datasets, each grounded in real Nepali financial data:

#	Source Name	Description	Rows	Domain
1	Key Financial Indicators of Commercial Banks	Quarterly financial metrics (solvency, capital, liquidity, etc.) for Nepali commercial banks — sourced from Nepal Rastra Bank	7,421	Nepalese Commercial Banking & Financial Indicators
2	BFI Nepal — Financial Institutions Registry	Branch-level facts for banks and financial institutions registered with Nepal Rastra Bank (NRB), including branch codes, districts, addresses, and establishment years	3,979	Banking
3	Saving Accounts Statistics by Province and District	Province and district-level savings account statistics across Nepal	585	Banking & Financial Services
4	NEPSE Listed Companies	Share listing data for companies listed on the Nepal Stock Exchange (NEPSE) — listed shares, paid-up value, and total paid-up capital	436	Finance / Capital Markets

---

Behavior Diversity

Each sample is associated with a behavior tag describing what kind of reasoning or action the model is expected to perform.

Behavior	Count	Description
grounded_single_metric_qa	7,421	Extract and state a single numeric financial metric from grounded data (e.g., a specific quarterly indicator for a specific bank)
select correct option	3,979	Multiple-choice question — select the single correct option from four choices
grounded_qa	585	Open-ended factual QA grounded in real statistics (e.g., savings data by province/district)
grounded_qa (NEPSE)	436	Open-ended factual QA grounded in NEPSE listing data (shares, paid-up capital)
Total	12,421	Rows with explicit behavior metadata

Note: The remaining 11,434 rows contain valid QA pairs with partial metadata; their behavior can be inferred from conversation content.

---

Question Pattern Diversity

By Question Type

Question Type	Count	Example Pattern
Multiple-choice grounded question	3,979	[statement]? ka. [A] kha. [B] ga. [C] gha. [D]
Grounded single-metric QA	7,421	[period] ma [bank] ko [metric] kati thiyo?
Grounded open QA	585	[region] ko [statistic] kati chha?
NEPSE factual QA	436	[company] ka kati wata seyarharu suchikrit chhan?

By Sub-Domain (BFI Branch Facts)

Sub-Domain	Count	What It Tests
district	928	Which district a bank branch is located in
open_year	923	Year a branch was established
code	911	Official branch code of a financial institution
type	680	Regulatory type/class of a financial institution
address	537	Registered address of a branch

By Financial Metric (Commercial Banking QA)

The grounded_single_metric_qa rows cover a wide range of quarterly financial indicators reported by Nepali commercial banks:

Grounding Field	Description
bank	Name of the commercial bank
period	Nepali fiscal period (e.g., 2077(12), 2078(3))
metric	Financial indicator (e.g., solvency core capital, liquidity ratio, NPL ratio)
value	Numeric value of the metric for that period

By NEPSE Data Fields

Grounding Field	Count	Description
Listed_Shares	143	Number of shares listed on NEPSE
Paidup_Value	149	Per-share paid-up value (NPR)
Total_Paidup	144	Total paid-up capital of the company
company	436	Full company name
symbol	436	NEPSE ticker symbol

---

Dataset Statistics

Conversation Length

Metric	Human Message	Assistant Message
Minimum length (chars)	28	8
Maximum length (chars)	334	327
Average length (chars)	96	75

Turn Distribution

Turns per Sample	Count
2 (1 human + 1 assistant)	23,855 (100%)

All samples follow a strict single-round instruction-following format.

---

Schema

Each row follows the ShareGPT format with an extended source_provenance field:

{
  "id": "string",
  "conversations": [
    {"from": "human", "value": "string"},
    {"from": "gpt",   "value": "string"}
  ],
  "source_provenance": {
    "id":              "string",
    "source":          "string",
    "source_name":     "string",
    "source_repo":     "string",
    "source_config":   "string",
    "source_split":    "string",
    "source_revision": "string",
    "source_row_id":   "string",
    "language":        "ne",
    "language_code":   "nep",
    "script":          "string",
    "license":         "Apache-2.0",
    "license_tier":    "string",
    "task_type":       "string",
    "generation_type": "real",
    "condition":       "real",
    "url":             "string",
    "metadata_json":   "string (JSON-encoded extra metadata)",
    "dataset_name":    "string",
    "dataset_split":   "string",
    "row_index":       0
  },
  "input_format":   "sharegpt",
  "source_task_id": "string"
}

Field Descriptions

Field	Description
id	Unique row identifier
conversations	List of turns: always [{human}, {gpt}]
source_provenance.source_name	Name of the originating dataset
source_provenance.language	Language code: ne (Nepali)
source_provenance.language_code	ISO 639-3 code: nep
source_provenance.license	Apache-2.0
source_provenance.generation_type	real — data grounded in real financial records
source_provenance.condition	real
source_provenance.metadata_json	JSON string with domain-specific metadata (grounding data, model name, dataset details)
input_format	Always sharegpt
source_task_id	Task-level identifier linking back to the generation pipeline

---

Sample Examples

1. Multiple-Choice — BFI Branch Facts

Human:
  11017198 kod bhaeko ra sundhara shakha naam garieko siddhartha bainka limitedko
  shakha kuna jillama awasthit chha?
  ka. lalitapur  kha. parwat  ga. parsa  gha. dang

Assistant:
  ka. lalitapur

Behavior: select correct option Sub-Domain: district Source: BFI Nepal Financial Institutions Registry

---

2. Grounded Single-Metric QA — Commercial Bank Indicators

Human:
  2077(12) ma nepal bainka limitedka lagi riporta garieko solbhensi kor kyapital kati thiyo?

Assistant:
  2077(12) ma nepal bainka limitedka lagi riporta garieko solbhensi kor kyapital 23220 thiyo.

Behavior: grounded_single_metric_qa Grounding: bank=Nepal Bank Limited, period=2077(12), metric=Solvency Core Capital Source: Key Financial Indicators of Commercial Banks (NRB Quarterly)

---

3. Grounded Open QA — Savings Statistics

Human:
  pradesh-starko rashtrawyapi samagrako kul janasankhya kati chha?

Assistant:
  rashtrawyapi pradesh-starko samagrako kul janasankhya 22,183,283 raheko chha.

Behavior: grounded_qa Grounding: level=province, region=Rashtrawyapi Source: Saving Accounts Statistics by Province and District in Nepal

---

4. NEPSE Listing QA

Human:
  egrikalchar debhalapamenta bainka limited (adbl) ka kati wata seyarharu suchikrit chhan?

Assistant:
  egrikalchar debhalapamenta bainka limited (adbl) ko rekarda garieko suchikrit seyarharuko sankhya 134,516,741 chha.

Behavior: grounded_qa Grounding: company=Agriculture Development Bank Limited, symbol=ADBL, Listed_Shares=134,516,741 Source: NEPSE Listed Companies

---

How to Load the Dataset

Using Hugging Face datasets library

from datasets import load_dataset

# Load the full dataset
ds = load_dataset("your-username/romanized-finance-dataset")
print(ds)
# DatasetDict({
#     train: Dataset({
#         features: ['id', 'conversations', 'source_provenance', 'input_format', 'source_task_id'],
#         num_rows: 23855
#     })
# })

# Access a single sample
sample = ds["train"][0]
print(sample["conversations"])
# [{'from': 'human', 'value': '...'}, {'from': 'gpt', 'value': '...'}]

Filter by Source / Behavior

from datasets import load_dataset
import json

ds = load_dataset("your-username/romanized-finance-dataset")["train"]

# Filter only BFI branch facts (MCQ samples)
bfi_samples = ds.filter(
    lambda x: x["source_provenance"]["source_name"] == "bfi_nepal_financial_institutions"
)
print(f"BFI samples: {len(bfi_samples)}")  # 3979

# Filter grounded metric QA samples
def is_grounded_metric(row):
    meta_str = row["source_provenance"].get("metadata_json", "")
    try:
        meta = json.loads(meta_str)
        return meta.get("behavior") == "grounded_single_metric_qa"
    except Exception:
        return False

metric_samples = ds.filter(is_grounded_metric)
print(f"Grounded metric QA samples: {len(metric_samples)}")  # 7421

Convert to Prompt-Completion Format

from datasets import load_dataset

ds = load_dataset("your-username/romanized-finance-dataset")["train"]

def to_prompt_completion(sample):
    convs = sample["conversations"]
    human = next((c["value"] for c in convs if c["from"] == "human"), "")
    gpt   = next((c["value"] for c in convs if c["from"] == "gpt"),   "")
    return {"prompt": human, "completion": gpt}

formatted = ds.map(to_prompt_completion, remove_columns=ds.column_names)
print(formatted[0])
# {'prompt': '...romanized nepali question...', 'completion': '...answer...'}

Load Directly from Local JSONL

from datasets import load_dataset

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

---

Models Used for Generation

Model	Rows	Task
gemini-3.5-flash-lite + gemini-3.6-flash	7,421	Commercial bank metric QA generation
gemini-3.5-flash-lite	436	NEPSE listing QA generation
Not applicable (template-based)	3,979	BFI branch MCQ (rule-based from structured data)
Not specified	585	Savings statistics QA

---

Intended Uses

Use Case	Suitability
Fine-tuning LLMs on Romanized Nepali financial text	✅ Primary use case
Instruction tuning / SFT (Supervised Fine-Tuning)	✅ ShareGPT format ready
Benchmarking Nepali financial QA	✅ Grounded, verifiable answers
Training MCQ models for Nepali finance exams	✅ 3,979 MCQ samples
Cross-lingual transfer learning (Nepali ↔ English)	✅ Romanized script bridges the gap
RAG evaluation on Nepali financial data	✅ Grounded provenance metadata included

---

Limitations

• All text is in Romanized Nepali (Nepali written in Latin script), not Devanagari. Models trained only on Devanagari Nepali may not directly benefit without transliteration.
• Financial data reflects a specific time window. Metrics from NRB quarterly reports are tied to their respective Nepali fiscal periods (e.g., 2077, 2078 BS).
• NEPSE listing data was sourced as of September 2024; share counts and paid-up values may have changed.
• BFI registry data reflects branches registered up to the dataset's collection date.
• Some rows (11,434) have partial metadata in source_provenance; the metadata_json field carries the full provenance detail for these.

---

License

This dataset is released under the Apache License 2.0.

The underlying source data is derived from:

• Nepal Rastra Bank (NRB) public financial data
• NEPSE public listing data (via Kaggle, license: RAIL)
• Open government financial statistics

See individual source_provenance.license and metadata_json fields per row for granular licensing details.

---

Citation

If you use this dataset in your research or projects, please cite:

@dataset{romanized_nepali_finance_2024,
  title     = {Romanized Nepali Finance Dataset},
  author    = {SABIN BHATTARAI},
  year      = {2026},
  language  = {Nepali (Romanized)},
  license   = {Apache-2.0},
  note      = {Instruction-following dataset covering Nepal's financial ecosystem
               in Romanized Nepali script. Sources include NRB commercial bank
               indicators, BFI registry, NEPSE listings, and savings statistics.}
}

Romanized Nepali Financial & Macroeconomic QA Datasets

A combined README for the five romanized (Latin-script) versions of the Nepali QA datasets built from Nepali company and central bank publications. Each file is a two-turn ShareGPT-style conversation (human question, GPT answer) with metadata.

The romanized files are the same records as the Devanagari originals, with the question and answer text written in Latin letters. Ids, sources, row order, and numeric values are preserved. This README covers all five files in one place: what they contain, how they relate to the originals, how to load them, and a diversity analysis for each file.

---

1. Contents

File	Rows	Source document	Domain	Size
gibl_2082_83_sharegpt_3k_romanized.jsonl	3,000	Global IME Bank Ltd., Annual Report FY 2082/83	Finance & Banking	4.56 MB
nrb_macroeconomic_2083_saun_3000_qa_romanized.jsonl	3,000	Nepal Rastra Bank, monthly macroeconomic report, Saun 2083	Macroeconomics & Finance	5.92 MB
barun_hydro_q3_2077_78_sharegpt_500__1__romanized.jsonl	500	Barun Hydropower Co. Ltd., unaudited Q3 results, FY 2077/78	Finance & Corporate Reporting	0.81 MB
kbl_sharesansar_sharegpt_200_romanized.jsonl	200	Kumari Bank Ltd. company profile and market data (ShareSansar)	Finance & Capital Market	0.29 MB
nbl_in_brief_sharegpt_200_romanized.jsonl	200	Nepal Bank Ltd., "NBL in Brief" web page	Finance & Banking	0.29 MB

Total: 6,900 records.

Companion files in this folder:

• romanized_diversity.py: computes all metrics in this README.
• romanized_diversity.json: the output of that script for the five files.

---

2. What "romanized" means here

Question and answer text is written in Latin letters using an informal transliteration of Nepali, for example:

Devanagari (original)	Romanized	Meaning
बैंक	bainka	bank
प्रतिशत	pratishat	percent
करोड	karod	crore (10 million)
छ / रहेको छ	chha / raheko chha	is / has been
को	ko	of
विस्तार	wistar	expansion
तथ्याङ्क	tathyanka	data
स्रोत	srot	source

The transliteration is not a formal standard. Spelling is fairly consistent within a file. In the GIBL file, the words checked for variant spellings (pratishat, bainka, aarthik, nikshep, wriddhi) each appeared with a single spelling. English terms are kept but lowercased, for example stp, rpa, and digital first, so the romanized file does not preserve the original capitalisation of acronyms.

2.1 What is and is not romanized

Part of the record	Romanized?
conversations (question and answer text)	Yes
id, source, source_row_id, source_name	No (unchanged from the original)
metadata_json	Mostly no. It is a JSON string, and its behavior_definition, content_language, and content_script fields are still in Devanagari.
content_script in metadata	Still says देवनागरी (Devanagari), which does not match the romanized text
Digits	Arabic (0-9) in both versions
Numeric values	Same as the original in every checked answer

---

3. Relation to the Devanagari originals

Every romanized file was checked against its original:

Check	GIBL	NRB	Barun	KBL	NBL
Same row count	Yes	Yes	Yes	Yes	Yes
Same ids, in order	3,000 / 3,000	3,000 / 3,000	500 / 500	200 / 200	200 / 200
Same source URL	3,000 / 3,000	3,000 / 3,000	500 / 500	200 / 200	200 / 200
Same metadata keys	Yes	Yes	Yes	Yes	Yes
Same fact_key	n/a	3,000 / 3,000	n/a	n/a	n/a
Numbers preserved in answers	945 / 945	2,970 / 2,970	262 / 262	155 / 163	53 / 54

Notes on the numeric check:

• A number is counted as preserved if the same digits appear in the romanized answer. Thousand separators and decimal formatting may differ.
• KBL and NBL have a few answers where the romanized version does not contain the same number. These cases were not checked individually.

3.1 Length

Romanized text is longer than the Devanagari original, because Latin spelling takes more characters than Devanagari for the same word:

Dataset	Median question length, original → romanized	Median answer length, original → romanized
GIBL	58 → 68 characters	91 → 106 characters
NRB	72 → 83 characters	92 → 107 characters
Barun	41 → 48 characters	75 → 86 characters
KBL	38 → 43 characters	23 → 24 characters
NBL	50 → 56 characters	64 → 75 characters

3.2 Unique answers

Romanization can make distinct Devanagari answers identical. KBL drops from 162 unique answers in the original to 155 in the romanized file, which suggests a few answers differed only in numeral script or punctuation. The other four files keep the same unique-answer count as the original.

---

4. Schema

Each line is one JSON object, with the same fields as the original files:

Field	Notes
id, source_row_id	Unique id, e.g. nrb_macro_2083_saun_0001.
conversations	Two turns: {"from": "human", "value": ...}, then {"from": "gpt", "value": ...}.
source, url	Source document URL. In the NRB file both fields hold the URL. In the other four files url is empty.
source_name, source_repo	Document title and publisher. Unchanged from the original.
language, language_code, script	ne, npi, Deva. script says Devanagari even though the text is romanized.
condition, generation_type	real in all records.
task_type	instruction-following.
license, license_tier	CC0: Public Domain (with the NRB note in that file), permissive.
metadata_json	JSON string. Parse it with json.loads.

4.1 Fields inside metadata_json

Field	Value	Varies?
behavior	Provide factual answer derived only from the dataset (NRB: ... from the official NRB report)	No
behavior_definition	Nepali text, Devanagari: answer while keeping the original question's meaning and facts, in pure Nepali	No
question_type	fact-based	No
generation_domain, generation_category, generation_sub_domain	Per-file values	No
content_language, content_script	नेपाली, देवनागरी	No
english_content_allowed	False	No
fact_key (NRB only)	e.g. inflation_overall, forex_usd	Yes, 75 values
report_period (NRB only)	आर्थिक वर्ष २०८३/८४ को पहिलो महिना (२०८३ साउन)	No

Across all five files, the metadata gives little variation within a file. The only field that varies is fact_key in the NRB file. Behaviour and question-pattern variety has to be measured from the text, as in the diversity sections below.

4.2 Example record (NRB, shortened)

{
  "id": "nrb_macro_2083_saun_3000",
  "conversations": [
    {"from": "human", "value": "aarthik barsha 2083/84 ko pahilo mahinako tathyankama aadharit dhitopatra bajar punjikaran kati raheko chha? (prashna nan. 5)"},
    {"from": "gpt", "value": "samiksha awadhima dhitopatra bajar punjikaran ru.4547 arba 75 karod raheko chha. (srot: nepal rashtra bainka)"}
  ],
  "metadata_json": "{\"fact_key\": \"...\", \"behavior\": \"Provide factual answer derived only from the official NRB report\", \"...\": \"...\"}"
}

---

5. How to load

Open files with encoding="utf-8". Parse metadata_json as a string.

5.1 Plain Python

import json

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

def to_qa(record):
    human = next(c["value"] for c in record["conversations"] if c["from"] == "human")
    gpt   = next(c["value"] for c in record["conversations"] if c["from"] == "gpt")
    meta  = json.loads(record["metadata_json"])   # a string in these files
    return {
        "id": record["id"],
        "question": human,
        "answer": gpt,
        "source": record["source"],
        "behavior": meta["behavior"],
        "fact_key": meta.get("fact_key"),          # only present in the NRB file
    }

rows = [to_qa(r) for r in read_jsonl("nrb_macroeconomic_2083_saun_3000_qa_romanized.jsonl")]
print(len(rows))   # 3000

5.2 All five files into one pandas DataFrame

import glob, os
import pandas as pd

frames = []
for path in glob.glob("*_romanized.jsonl"):
    df = pd.DataFrame([to_qa(r) for r in read_jsonl(path)])
    df["file"] = os.path.basename(path)
    frames.append(df)

all_qa = pd.concat(frames, ignore_index=True)
print(all_qa.groupby("file").size())

5.3 Matching a romanized record to its original

Because ids are identical, you can join the romanized file to the Devanagari original by id:

orig = {r["id"]: r for r in read_jsonl("nrb_macroeconomic_2083_saun_3000_qa.jsonl")}
for r in read_jsonl("nrb_macroeconomic_2083_saun_3000_qa_romanized.jsonl"):
    original = orig[r["id"]]
    # original["conversations"] is the Devanagari text, r["conversations"] is the romanized text

5.4 Cleaning the known artifacts

Two artifacts should be cleaned before use (see Section 9):

import re

def clean(text):
    text = text.replace("\u200d", "")                    # zero-width joiner left from Devanagari conjuncts
    text = re.sub(r"\bru\.(\d)", r"ru. \1", text)        # restore the space in "ru.41" -> "ru. 41"
    return text

5.5 Converting to chat format

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

5.6 Shared helper scripts

The scripts in the scripts/ folder of the earlier dataset bundle also read these files, because they handle metadata_json as a string. For the romanized diversity metrics, use this folder's script:

python romanized_diversity.py --data_dir /path/to/files --originals_dir /path/to/originals --out romanized_diversity.json

---

6. Dataset overview

Metric	Barun	GIBL	KBL	NBL	NRB
Records	500	3,000	200	200	3,000
Unique questions	499	2,964	199	200	2,995
Unique answers	288	327	155	98	1,895
Unique answers (% of records)	57.6%	10.9%	77.5%	49.0%	63.2%
Unique questions after removing prefixes	384	1,654	196	145	2,793
Most repeated base question (count)	2	25	2	3	2
Median question length (chars)	48	68	43	56	83
Median answer length (chars)	86	106	24	75	107
Rows that are fully ASCII	99.4%	91.2%	100.0%	98.5%	100.0%
Answer words shared with question (avg)	0.22	0.23	0.07	0.11	0.37

---

7. Question and answer diversity

The Devanagari analysis used keyword rules for Devanagari. These romanized files use the matching rules for romanized keywords (kati, kasari, kina, kasto, and so on). Percentages are close to, but not identical to, the Devanagari analysis, so compare within this document rather than against the earlier one.

7.1 Question types

Question type	Barun	GIBL	KBL	NBL	NRB
Quantity (kati / kun parimaan)	29.6%	19.3%	39.5%	10.5%	54.2%
What (ke)	22.0%	9.9%	28.5%	49.0%	11.7%
Quality / description (kasto)	22.2%	12.0%	3.5%	3.0%	13.8%
Method (kasari)	1.8%	30.7%	0.0%	10.0%	0.0%
Reason (kina)	2.4%	2.7%	0.0%	5.0%	0.0%
Who / identity (ko + ho / hun)	0.0%	2.6%	0.0%	0.0%	0.0%
Which (kun)	0.0%	0.0%	0.0%	0.0%	0.0%
Location (kahan)	4.8%	0.9%	3.0%	0.0%	0.0%
Time (kahile / kun miti)	0.0%	1.5%	2.5%	1.0%	0.0%
Other (yes/no, imperative, unmatched)	17.2%	20.5%	23.0%	21.5%	20.3%

Notes:

• The "which" and "who" rules are narrow. Possessive "ko" (of) is excluded on purpose, which keeps the "who" count low.
• The "other" share is higher than in the Devanagari analysis, because the romanized keyword list is shorter. Treat it as an unmatched remainder.

7.2 Answer types

Answer type	Barun	GIBL	KBL	NBL	NRB
Numeric, inside a sentence	48.2%	31.2%	24.5%	13.5%	98.7%
Descriptive / short text (no number)	46.0%	68.5%	18.5%	73.0%	1.0%
Numeric, short	4.2%	0.3%	56.5%	13.5%	0.3%
Not in document / refusal	1.6%	0.0%	0.5%	0.0%	0.0%

The KBL answers are mostly bare values (56.5% short numeric), the GIBL and NBL answers are mostly descriptive, and the NRB answers almost all contain a figure.

7.3 Instruction prefixes

Dataset	Share of questions with a prefix	Distinct prefixes	Most common prefix
Barun	23.2%	8	"please tell me" / "please give information" / "brief" / "detailed" / "I want to know" (19 each)
GIBL	44.2%	7	"brief" (199), "please tell me" (197), "details" (192)
KBL	1.5%	3	(1 each)
NBL	27.5%	5	(11 each)
NRB	7.0%	14	"latest figure" (20), "please give information" (19)

Prefix examples, with translations:

• kripaya bataunuhos: (please tell)
• janakari dinuhos: (please give information)
• sankshepma bhannuhos: (say briefly)
• wistarma bhannuhos: (say in detail)
• ma janna chahanchhu, (I want to know)
• pachhillo aankada: (latest figure)

7.4 Behaviour patterns

Pattern	Where it appears
Direct value lookup	KBL (56.5% short numeric), Barun
Descriptive sentence	GIBL (68.5%), NBL (73.0%)
Method questions ("kasari")	GIBL (30.7%), NBL (10.0%)
Hypothetical framing ("yadi" = if)	GIBL only, 178 records
Corporate disclaimer or assurance	Barun (7 answers contain "srot", plus assurance answers)
Source citation in the answer	NRB: 591 records contain "(srot: nepal rashtra bainka)"
Prashna number tag, e.g. "(prashna nan. 5)"	NRB: 705 questions
Refusal / "not in the document"	Barun 8, KBL 1, others 0

7.5 Answer reuse and ambiguity

Metric	Barun	GIBL	KBL	NBL	NRB
Answers reached by more than one distinct question	96	327	28	43	652
Base questions with more than one distinct answer	1	11	0	0	173
Most reused answer (count)	4	12	5	4	5

Notes:

• In GIBL, the 11 questions with more than one answer include the question on who the director is, which has several valid answers. These should be reviewed before training.
• In NRB, the 173 "ambiguous" base questions are mostly paraphrase variants of the same figure. They are not contradictions. The fact-level consistency check described in the NRB README showed that the dominant figure matches across most records in each fact.

---

8. Format and numeric details

Metric	Barun	GIBL	KBL	NBL	NRB
Answers with comma-grouped numbers (e.g. 3,693,387)	136	196	26	0	34
Answers with decimal numbers	78	428	98	14	1,710
Answers with "ru." before a number	ru. with space: 195	222	23	17	0 (see 9.1)
Answers with "ru." and no space before a number	0	0	0	0	1,078
Answers with "srot" (source)	7	25	0	0	591

Number formatting is not uniform. GIBL and Barun mix comma-grouped integers with decimals. NRB uses decimals almost throughout. This is useful variety for number normalisation, but exact-match evaluation will need a normaliser.

---

9. Known artifacts and quality issues

1. Missing space after "ru." in NRB (1,078 answers). The romanized NRB answers contain strings such as ru.41 arba 89 karod. The original has रु.४१ अर्ब, so the space was lost in conversion. The other romanized files keep the space (ru. 41). Apply the cleaning step in Section 5.4 before use.
2. Zero-width joiner in GIBL (215 records) and NBL (3 records). The character \u200d is left over from Devanagari conjuncts, as in pur\u200dyaeko (puryaeko). It breaks tokens and string matching. Remove it with the cleaning step in Section 5.4.
3. Division sign in GIBL (88 records). The character ÷ appears in ratio labels such as kul karja ÷ nikshep anupat, from the original. It is correct but non-standard for Latin text.
4. Unicode symbols in Barun. The arrow → (10 times) and the minus sign − (once) are kept from the original.
5. Metadata still in Devanagari. behavior_definition, content_language, content_script, and report_period are not romanized. content_script says देवनागरी even for the romanized text.
6. Keyword rules are romanized but still approximate. Question-type shares come from keyword matching and are not human labels.
7. Repetition in GIBL. Only 327 unique answers cover 3,000 records. This is the same as in the Devanagari file and needs deduplication before training.
8. Templated attribution in NRB. The source tag (srot: nepal rashtra bainka) appears in 591 answers, and phrases such as "samiksha awadhima" (in the review period) repeat heavily.
9. Snapshot data. NRB answers describe one month (Saun 2083). KBL answers are market values with no date. Both must be dated before use for current-value questions.

---

10. Limitations

1. Romanization is informal. There is no standard scheme, so the same Nepali word may be written in different ways in other sources. Within these files the key words are consistent, but the scheme should be documented for any new data added.
2. Metadata is largely uniform. Only the NRB file has a varying field (fact_key). The others carry one behaviour and one question type for all records.
3. Few facts in NRB. 75 fact_key values cover 3,000 records. Random splits will overstate generalisation; split by fact_key instead.
4. No refusals in GIBL, NBL, or NRB. The model is not shown what to answer when information is not in the document.
5. No forward-looking questions in NRB. Forecast and outlook questions are absent.
6. Single-turn only. All records are one question and one answer.
7. Licence. The metadata declares CC0 public domain. The underlying documents belong to the companies and to NRB, and their own terms may differ. Confirm before redistribution.
8. Factual accuracy not verified. Figures were not checked against the source PDFs or web pages in this README.
9. Small files. KBL and NBL have 200 records each, so their percentages are noisy: one record moves a share by 0.5 points.

---

11. Recommended steps

Priority	Step	Reason
High	Apply the cleaning step (Section 5.4) to all files: remove \u200d, restore the space after ru..	Artifacts in Section 9.
High	Add refusal examples ("not in the report") to GIBL, NBL, and NRB.	Section 10, item 4.
High	Resolve the 11 ambiguous GIBL questions, e.g. by naming a specific director.	Conflicting targets.
High	Deduplicate GIBL answers and cap repeats per answer.	327 unique answers for 3,000 records.
Medium	Romanize content_script and behavior_definition, or document that they stay in Devanagari.	Metadata/text mismatch.
Medium	Add the reporting period to every NRB answer.	Snapshot data.
Medium	Add a snapshot date to KBL records.	Market values change daily.
Low	Add multi-turn follow-ups.	All records are single-turn.

---

12. Files

romanized_datasets/
├── README.md                    This file
├── romanized_diversity.py       Diversity and alignment metrics for all five romanized files
└── romanized_diversity.json     Output of romanized_diversity.py for the five files

The five romanized .jsonl files are not copied into this folder. Pass their folder with --data_dir when running the script.

Regenerate the metrics:

python romanized_diversity.py --data_dir /path/to/files --originals_dir /path/to/originals --out romanized_diversity.json

NRB Annual Report 2024/25 Q&A Dataset (Romanized)

The romanized (Latin-script) version of the NRB Annual Report 2024/25 Q&A dataset. The file has 3,000 single-turn records in ShareGPT format. The question and answer text is written in Latin letters using informal transliteration of Nepali. Ids, sources, metadata, fact keys, and numeric values match the Devanagari original (nrb_annual_report_2024_25_nepali_3000_qa.jsonl).

For the romanization scheme and the general approach, see the combined README for the romanized datasets. This README covers only this file.

---

1. Contents

File	Rows	Size	Source	Devanagari original
nrb_annual_report_2024_25_nepali_3000_qa_romanized.jsonl	3,000	5.26 MB	NRB Annual Report 2024/25	nrb_annual_report_2024_25_nepali_3000_qa.jsonl (6.12 MB)

The source_name field says Nepali, but the source and url fields point to Annual-Report-2024-25-English.pdf. This mismatch is inherited from the original. Confirm which document the questions came from before citing the dataset.

---

2. Relation to the Devanagari original

Check	Result
Row count	3,000 / 3,000
Same ids, in order	3,000 / 3,000
Same source URL	3,000 / 3,000
Same top-level keys	Yes, all records
Same metadata keys	Yes, all records
Same fact_key	3,000 / 3,000 (49 distinct values)
Numbers in questions preserved	3,000 / 3,000
Numbers in answers preserved	3,000 / 3,000
Devanagari characters left in text	0
Non-ASCII characters in text	0 (every question and answer is pure ASCII)

Numbers are checked after converting Devanagari digits to ASCII in the original. Thousand separators and decimal formatting are the same in both files, since the original never used thousand separators.

2.1 Length

Field	Original (median chars)	Romanized (median chars)	Ratio
Question	80	90	1.14
Answer	103	116	1.13

2.2 What is romanized and what is not

Part of the record	Romanized?
conversations (question and answer text)	Yes
id, source, url, source_name, source_repo	No (unchanged)
metadata_json	Mostly no. It is a JSON string. behavior is in English. behavior_definition, content_language, and content_script stay in Devanagari.
content_script	Still says देवनागरी, which does not match the romanized text
Digits	ASCII (0–9)

2.3 Transliteration in this file

Spelling is consistent across the file. Key words appear with a single spelling each, and no alternate spellings were found for the words checked:

Word	Count	Alternate spellings found
pratishat (percent)	1,603	0
pratiwedana (report)	1,768	0 (prativedan: 0)
arba (billion)	1,519	0
karod (crore)	0	Not used in this file; see Section 6
bainka (bank)	273	0 (banka: 0)
samiksha (review)	567	0
aarthik (fiscal)	556	0

The word karod does not appear in this file, because the annual report uses amounts in arba (billion) and did not need the crore unit in the checked answers.

---

3. Schema

Field	Value / meaning
id, source_row_id	Unique id, nrb_annual_2024_25_ne_0001 to nrb_annual_2024_25_ne_3000.
conversations	Two turns: human question, then GPT answer.
source, url	Same value in both: the English annual report PDF URL.
source_name	nrb_annual_report_fy_2024_25_nepali
source_repo	Nepal Rastra Bank - Economic Research Department
language, language_code, script	ne, npi, Deva.
condition, generation_type	real in all records.
task_type	instruction-following.
license, license_tier	CC0: Public Domain (based on official NRB public report), permissive.
metadata_json	JSON string. Parse with json.loads.

3.1 Fields inside metadata_json

Field	Value	Varies?
behavior	Provide factual answer derived only from the official NRB Annual Report 2024/25	No
behavior_definition	Nepali text (Devanagari)	No
question_type	fact-based	No
generation_domain	Macroeconomics & Finance	No
generation_category	NRB Annual Report	No
generation_sub_domain	Fiscal Year 2024/25	No
content_language, content_script	नेपाली, देवनागरी	No
english_content_allowed	False	No
fact_key	Names the report figure, e.g. industry_share, gdp_basic, vat. 49 distinct values.	Yes
report_period	आर्थिक वर्ष २०२४/२५ (Devanagari)	No

3.2 Example record

{
  "id": "nrb_annual_2024_25_ne_0002",
  "conversations": [
    {"from": "human", "value": "nepal rashtra bainkako warshik pratiwedana 2024/25 anusar kul garhasthya utpadanama audyogik kshetrako ansha kati raheko chha? (prashna 8)"},
    {"from": "gpt", "value": "kul garhasthya utpadanama audyogik kshetrako ansha 12.83 pratishat raheko chha. (srot: nepal rashtra bainkako warshik pratiwedana 2024/25)"}
  ],
  "metadata_json": "{\"fact_key\": \"industry_share\", \"behavior\": \"Provide factual answer derived only from the official NRB Annual Report 2024/25\", \"...\": \"...\"}"
}

---

4. How to load

import json

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

def to_qa(record):
    human = next(c["value"] for c in record["conversations"] if c["from"] == "human")
    gpt   = next(c["value"] for c in record["conversations"] if c["from"] == "gpt")
    meta  = json.loads(record["metadata_json"])        # a string in this file
    return {
        "id": record["id"],
        "question": human,
        "answer": gpt,
        "fact_key": meta["fact_key"],
    }

rows = [to_qa(r) for r in read_jsonl("nrb_annual_report_2024_25_nepali_3000_qa_romanized.jsonl")]
print(len(rows))   # 3000

Split by fact_key to keep questions about the same figure out of both train and test sets:

import random
from collections import defaultdict

by_fact = defaultdict(list)
for r in rows:
    by_fact[r["fact_key"]].append(r)

keys = sorted(by_fact)
random.Random(42).shuffle(keys)
test_keys = set(keys[:10])
train = [r for k in keys if k not in test_keys for r in by_fact[k]]
test  = [r for k in test_keys for r in by_fact[k]]

To match a romanized record to its Devanagari original, join on id:

orig = {r["id"]: r for r in read_jsonl("nrb_annual_report_2024_25_nepali_3000_qa.jsonl")}

4.1 Cleaning

Two items are worth fixing before use (see Section 6). The first is optional, and the second is a real issue:

import re

def clean(text):
    text = re.sub(r"\bru\.(\d)", r"ru. \1", text)     # optional: "ru.223.72" -> "ru. 223.72"
    text = text.replace("\u200d", "")                  # zero-width joiner, if present
    return text

Remove or move the question-number tags before evaluation if you do not want the model to learn them:

def strip_tags(text):
    text = re.sub(r"\s*\(srot[^)]*\)", "", text, flags=re.I)        # source citation
    text = re.sub(r"\s*\(prashna\s*(nan\.?\s*)?\d+\)", "", text, flags=re.I)  # question-number tag
    return text.strip()

---

5. Diversity

The diversity metrics below come from romanized_diversity.py (in romanized_datasets/), which uses keyword rules for romanized Nepali. Compare against the other romanized files, or against the Devanagari analysis, with care: the keyword lists differ, so percentages are close but not identical.

5.1 Overview

Metric	Value
Records	3,000
Unique questions	2,907 (96.9%)
Unique answers	1,702 (56.7%)
Unique questions after removing instruction prefixes	1,587
Distinct fact_key values	49
Records per fact (min / median / max)	20 / 61 / 92
Average answer word overlap with question	0.39
Exact duplicate question pairs	93 (186 records)

The unique-answer count and the duplicate count match the Devanagari original exactly, so romanization did not merge any answers in this file.

5.2 Question types (romanized keyword rules)

Question type	Share
Quantity (kati / kun parimaan)	49.1%
Other (yes/no, imperative, unmatched)	27.6%
What (ke)	17.4%
Quality / description (kasto)	5.9%

The keyword rules found no "how" (kasari), "why" (kina), "when" (kahile), "where" (kahan), or "which" (kun) questions. The question set is dominated by quantity, what, and quality questions, which limits the explanatory use of this file.

5.3 Instruction prefixes

47.1% of questions (1,413 records) start with an instruction prefix. The most common prefixes, in romanized form and translation:

Prefix (romanized)	Translation	Records
janakari dinuhos	please provide information	123
samiksha awadhiko aadharma	based on the review period	110
warshik pratiwedanama	in the annual report	110
tathyanka ke bhanchha	what does the data say	107
ke tapailai thaha chha	do you know	106

Fifteen distinct prefixes appear in the top list. Most differ only in wording.

5.4 Answer format

Answer type	Share
Number inside a sentence	100%
Descriptive sentence without a number	0%

Every answer states a figure. There are no purely descriptive answers, which limits variety in answer shape.

5.5 Citations and tags

Pattern	Records	Share
Answer contains the citation (srot: ...)	1,011	33.7%
Question contains the question-number tag (prashna N)	736	24.5%

The question-number tag runs from 1 to 8. Each value appears about 90 times (8 appears 96 times, 1 appears 94 times, 6 appears 93 times, 4 appears 92 times). It says nothing about the content of the question, but a model can learn it as a cue. The tag appears in the same form as the Devanagari original, (प्रश्न N), translated into (prashna N).

5.6 Duplicates and ambiguity

• Exact duplicate questions: 93 pairs. In every pair, the two answers give the same figures. Each pair differs only in wording or in the citation and review-period phrasing. There are no duplicates with conflicting numbers.
• Base questions with more than one answer: 1,148 by the shared script's count. These are paraphrase variants of the same figure, not contradictions.
• Answers reached by more than one question: 705.

5.7 Fact-level consistency

For each fact_key, the first number in each answer is the main figure. The check was run on the Devanagari original and gave the same result here: in 40 of 49 facts a second leading number also appears, usually as a minority value (for example, travel_expense gives 223.72 in 52 of 60 answers and 18.1 in 8). These minority values are most likely other numbers in the same answer rather than conflicts. The check is automated and approximate.

---

6. Known artifacts

1. No space after ru. (1,481 answers, 49.4%). The romanized answers contain strings such as ru.223.72 arba. The Devanagari original has the same style (रु.२२३.७२ अर्ब, no space), so romanization did not introduce this. It is a style inconsistency in the source data, since other datasets in this collection write ru. 41 with a space. Normalise it if you want consistent spacing (Section 4.1). It is not required for correctness.
2. Question-number tags. 736 questions (24.5%) carry (prashna N). This matches the original and is a generation artifact.
3. Source mismatch. source_name says Nepali, but source and url point to the English PDF. This is inherited from the original.
4. Metadata still in Devanagari. behavior_definition, content_language, content_script, and report_period are not romanized. content_script says देवनागरी for romanized text.
5. Duplicate questions. 93 pairs (186 records), as in the original.

No zero-width joiners or other non-ASCII characters were found in this file.

---

7. Limitations

1. Informal romanization. There is no standard scheme. The spelling is consistent within this file, but it may not match other romanized sources.
2. Few facts. 49 facts cover 3,000 records. Split by fact_key to measure generalisation.
3. Metadata is uniform. Only fact_key varies. The behaviour and question type are the same for all records.
4. No refusals. There are no answers stating that information is not in the report.
5. No comparisons or forecasts in the question text. The file does not show how to answer change-over-time or outlook questions.
6. No explanatory questions. The keyword rules found no "how", "why", or "where" questions.
7. Single-turn only.
8. Heuristic labels. Question-type shares come from keyword rules, not human annotation.
9. Factual accuracy not verified. Figures were not checked against the report in this README.
10. Licence. The metadata declares CC0 based on the official NRB report. Check NRB's terms of use before redistribution.

---

8. Recommended steps

Priority	Step	Reason
Medium	Normalise the ru. spacing (Section 4.1) for consistency across datasets.	Section 6, item 1.
High	Remove or move the (prashna N) tags.	Section 5.5.
High	Confirm the source document and correct source / url or source_name.	Section 6, item 3.
High	Deduplicate the 93 duplicate question pairs.	Section 5.6.
High	Add refusal examples such as "not in the annual report" with the correct answer.	Section 7, item 4.
Medium	Add explanatory "why" and "how" questions.	Section 7, item 6.
Medium	Romanize content_script and behavior_definition, or document that they stay in Devanagari.	Section 6, item 4.
Low	Add multi-turn follow-ups.	All records are single-turn.

---

9. Files

nrb_annual_report_2024_25_romanized/
├── README.md                  This file
└── romanized_diversity.json   Diversity metrics for this file, with alignment against the Devanagari original

The data file nrb_annual_report_2024_25_nepali_3000_qa_romanized.jsonl is not copied into this folder. To regenerate the metrics:

python ../romanized_datasets/romanized_diversity.py --data_dir /path/to/files --originals_dir /path/to/originals --out romanized_diversity.json

The script matches each *_romanized.jsonl file in --data_dir to a Devanagari original of the same name without _romanized.

Nepali Factual QA Datasets (Romanized and Devanagari)

Three Nepali question-answering datasets built from public financial and corporate documents. Each record is a single-turn factual question with a short factual answer, stored as ShareGPT-style JSON Lines with per-row metadata.

Two of the three datasets exist in two written forms: Devanagari (original) and romanized (Latin script). This README covers both forms in one place, explains how they relate, and reports the verified quality and diversity of each.

---

---

1. Files at a Glance

File	Dataset	Script	Rows	Size	Status
bank_failures_ROMANIZED_qa.jsonl	Bank failures	Romanized	25,472	38.2 MB	Use this one
bank_failures_ROMANIZED_qa_1.jsonl	Bank failures	Romanized	25,472	38.2 MB	Byte-identical duplicate of the file above (same MD5). Do not load both.
sahas_urja_ipo_qa_final_unique.jsonl	Sahas Urja IPO	Devanagari	262	0.44 MB	Original
sahas_urja_ipo_qa_final_unique_romanized.jsonl	Sahas Urja IPO	Romanized	262	0.41 MB	Aligned 1:1 with the Devanagari file by id
men_ar_2081_82_sharegpt_2000.jsonl	Mountain Energy AR	Devanagari	2,000	3.25 MB	Original
men_ar_2081_82_sharegpt_2000_ROMANIZED.jsonl	Mountain Energy AR	Romanized	2,000	2.92 MB	Aligned 1:1 with the Devanagari file by id. Contains a known residual-character issue (Section 8, item 3).

Total unique QA pairs: 27,734 (bank 25,472; Sahas 262; Mountain Energy 2,000). Each Sahas and Mountain Energy QA pair exists twice, once per script, so the Devanagari and romanized versions of those two datasets contain the same facts.

The bank dataset has no Devanagari version.

---

2. Romanized vs Devanagari

2.1 What was checked

The romanized files were compared against their Devanagari counterparts. The checks below were run on the actual files.

Check	Sahas Urja	Mountain Energy	Result
Same row count	262 vs 262	2,000 vs 2,000	Pass
Same id order	Yes	Yes	Pass
Same source_row_id	Yes	Yes	Pass
Digit sequences preserved in answers (after converting Devanagari digits to Latin)	0 mismatches	0 mismatches	Pass
Devanagari characters left in romanized text	0	39 questions (4 of these also in answers)	Fail for Mountain Energy (Section 8)
Unique questions	262	2,000	Same as original
Unique answers	256	360	Same as original

2.2 Digit handling

Devanagari digits were converted to Latin digits in the romanized files. For example, ०१-४४१९१८२ became 01-4419182. The digit sequences match exactly across the two versions, so numeric facts are preserved.

2.3 Side-by-side examples

Devanagari question	Romanized question
साहस उर्जा लिमिटेड कहाँ अवस्थित छ?	sahas urja limited kaha awasthit chha?
माउन्टेन इनर्जी नेपाल लिमिटेडको अध्यक्ष को हुन्?	mauntena inarji nepal limitedko adhyaksha ko hun?

Devanagari answer	Romanized answer
हात्तिसार, काठमाण्डौं-१, नेपालमा अवस्थित छ।	hattisar, kathamandau-1, nepalma awasthit chha.
श्री प्रदिप जंग पाण्डे अध्यक्ष हुन्।	shri pradip janga pande adhyaksha hun.

2.4 Romanization style

The romanization is a custom Latin spelling, not a standard transliteration scheme. Key function words are spelled consistently across the files, for example chha (851 times in the romanized Sahas and Mountain Energy files), ho (959), kati (981), ru. (767), and thiyo (157). Proper nouns and technical terms are transliterated by sound, so they will not always match an English or official spelling (for example hattisar, kathamandau, mauntena inarji).

---

3. File Format and Schema

All six files are JSON Lines (one JSON object per line) and share the same top-level schema.

3.1 Top-level fields

Field	Type	Description
id	string	Unique row identifier, for example bank_fail_faq_0001, sahas_ipo_0001, men_ar_2081_82_0001. Identical across the Devanagari and romanized versions of the same row.
conversations	list of 2 objects	Turn 1 has from: "human" (the question). Turn 2 has from: "gpt" (the answer).
source	string	URL of the original document (Mountain Energy has this in source, while url is empty).
source_name	string	Human-readable name of the source document.
source_repo	string	Publisher or repository name.
source_config, source_split, source_revision	string	Always default, train, and v1 respectively.
source_row_id	string	Same as id.
language	string	Always ne (Nepali).
language_code	string	Always npi (ISO 639-3 code for Nepali).
script	string	Declared script. Always Deva in all six files, including the romanized ones. See Section 8.
license	string	Always Apache-2.0.
license_tier	string	Always permissive.
task_type	string	Always instruction-following.
generation_type, condition	string	Always real.
url	string	Source URL (empty for Mountain Energy).
metadata_json	string	A JSON-encoded string, not a nested object. Call json.loads on it.

3.2 Fields inside metadata_json

Key	Bank Failures	Sahas Urja IPO	Mountain Energy AR
generation_domain	Banking & Finance	धितोपत्र र ऊर्जा (Securities & Energy)	Finance & Hydropower
generation_category	Bank Failures	आईपीओ विवरणपत्र (IPO prospectus)	Annual Report
generation_sub_domain	FDIC Commercial Bank Failures 1934-Present	साहस उर्जा	Mountain Energy Nepal Ltd. 12th AR FY 2081/82
behavior	Provide factual answer derived only from the dataset	जानकारीमूलक उत्तर दिने (Give informational answer)	Provide factual answer derived only from the dataset
behavior_definition	Keep the meaning of the question and the facts of the answer while converting it into pure Nepali	Same	Same
question_type	fact-based	तथ्यमा आधारित प्रश्न (fact-based question)	fact-based
content_language	नेपाली	नेपाली	नेपाली
content_script	देवनागरी	देवनागरी	देवनागरी
english_content_allowed	False	False	False
original_row_index	Source row index (Section 4)	not present	not present
column_focus	Table column targeted by the QA pair (Section 4)	not present	not present

The content_script, content_language, and behavior_definition values describe a Devanagari target. These are the same in the romanized files, which is a metadata inconsistency (Section 8, item 2).

3.3 Example record (bank, shortened)

{
  "id": "bank_fail_faq_0001",
  "conversations": [
    {"from": "human", "value": "bank of america trust co. bainka kahile asaphal bhayo?"},
    {"from": "gpt",   "value": "bank of america trust co. bainka 4/19/1934 ma asaphal bhayo."}
  ],
  "language_code": "npi",
  "script": "Deva",
  "metadata_json": "{\"question_type\": \"fact-based\", \"column_focus\": \"Failure Date\", ...}"
}

---

4. Dataset 1: Bank Failures

4.1 What it contains

This dataset turns the FDIC commercial bank failure table (1934 onward) into question-answer pairs. Each source row describes one failed bank. There are 3,184 source rows, and 8 QA pairs are generated per row, one for each selected column. That gives 3,184 × 8 = 25,472 QA pairs.

The file is romanized only.

4.2 Generation method

Each source row is expanded into 8 questions, each targeting one column (column_focus). Each question uses a fixed Nepali template, and each answer fills a fixed answer template with the source value.

column_focus	Question template	Answer template	Answer type
Failure Date	{bank} kahile asaphal bhayo?	{bank} {date} ma asaphal bhayo.	Date (M/D/YYYY)
Headquarters	{bank} ko mukhyalaya sthana ke ho?	{bank} ko mukhyalaya {city, state} ma awasthit thiyo.	Place
Institution Type	{bank} kasto prakarko sanstha thiyo?	{bank} {type} prakarko sanstha thiyo.	Category
Charter Type	{bank} ko chartar prakar ke thiyo?	{bank} ko chartar prakar {type} thiyo.	Category
Insurance Fund	{bank} kuna bima kosh antargat thiyo?	{bank} {fund phrase} bima kosh antargat thiyo.	Mixed text
Transaction Type	{bank} ko asaphalata lenadena prakar ke thiyo?	{bank} ko lenadena prakar {type} thiyo.	Category
Total Deposits	{bank} ko kul nikshep rakam kati thiyo?	{bank} ko kul nikshep {number} thiyo.	Number
Total Assets	{bank} ko kul sampatti mulya kati thiyo?	{bank} ko kul sampatti {number} thiyo.	Number

4.3 Content statistics

Metric	Value
Source rows	3,184
QA pairs per source row	8 (all rows)
Approximate distinct bank names	3,183
Failure years covered	1934 to 2017
Answers containing a digit	9,432 (37.0%)
Median question length	9 words (min 5, max 16)
Median answer length	9 words (min 6, max 18)

Failure-date distribution by decade (3,184 failure-date answers):

Decade	Failures
1930s	288
1940s	95
1950s	26
1960s	38
1970s	67
1980s	1,322
1990s	846
2000s	187
2010s	315

The 1980s and 1990s dominate, which reflects the historical wave of US savings and loan and small-bank failures.

4.4 Categorical value distribution

Column	Value	Count
Charter Type	state	1,815
Charter Type	federal	714
Charter Type	federal/state	655
Transaction Type	acquisition	2,250
Transaction Type	payout	525
Transaction Type	transfer	371
Transaction Type	management	36
Transaction Type	privatization	2
Institution Type	commercial bank	2,212
Institution Type	savings association	655
Institution Type	savings bank	317

Insurance Fund answers are mixed text, not a clean label. Fund codes appear inside the answer: fdic (1,348 answers), rtc (704), bif (524), and saif (6). Some answers contain the bank's own legal name before the fund phrase, so the column needs text parsing before it can be used as a category.

Total Deposits and Total Assets are plain numbers with no unit stated in the data (for example 91 or 1064).

4.5 Example pairs (one per column)

column_focus	Question	Answer
Failure Date	bank of america trust co. bainka kahile asaphal bhayo?	bank of america trust co. bainka 4/19/1934 ma asaphal bhayo.
Headquarters	bank of america trust co. ko mukhyalaya sthana ke ho?	bank of america trust co. ko mukhyalaya pittsburgh, pa ma awasthit thiyo.
Institution Type	bank of america trust co. kasto prakarko sanstha thiyo?	bank of america trust co. commercial bank prakarko sanstha thiyo.
Charter Type	bank of america trust co. ko chartar prakar ke thiyo?	bank of america trust co. ko chartar prakar state thiyo.
Insurance Fund	bank of america trust co. kuna bima kosh antargat thiyo?	bank of america trust co. fdic bima kosh antargat thiyo.
Transaction Type	bank of america trust co. ko asaphalata lenadena prakar ke thiyo?	bank of america trust co. ko lenadena prakar payout thiyo.
Total Deposits	bank of america trust co. ko kul nikshep rakam kati thiyo?	bank of america trust co. ko kul nikshep 1064 thiyo.
Total Assets	bank of america trust co. ko kul sampatti mulya kati thiyo?	bank of america trust co. ko kul sampatti 1435 thiyo.

---

5. Dataset 2: Sahas Urja IPO Prospectus

5.1 What it contains

262 hand-curated, unique QA pairs extracted from the IPO prospectus of Sahas Urja Limited. Questions cover company identity, registration, capital and share issue, the hydropower project and its technical parameters, governance, financial figures, and the regulatory process.

Both forms share the same 262 ids. Every question is unique, and there are 256 unique answers.

5.2 Topic distribution

Topics were assigned with a keyword classifier on the Devanagari text. Because the romanized file is aligned by id, the same assignment applies to it. Counts are approximate because some questions match more than one topic and the classifier returns the first match.

Topic	Approximate count	Example (Devanagari)
Project and technical	47	जेनेरेटरको फ्रिक्वेन्सी र भोल्टेज कति हो?
Capital and share issue	42	कम्पनीको जारी पूँजी कति हो?
Company identity and contact	17	साहस उर्जा लिमिटेड कहाँ अवस्थित छ?
Financial and fiscal	13	आ.ब. ०७६/०७७ मा प्रति शेयर नेटवर्थ कति थियो?
Process and regulator	9	नेपाल धितोपत्र बोर्डबाट विवरणपत्र स्वीकृति कहिले प्राप्त भयो?
Registration and legal	6	कम्पनी कहिले दर्ता भएको थियो?
Governance and promoter	5	कम्पनीमा हाल कति जना संस्थापक शेयरधनी छन्?
General or other	123	—

5.3 Question style (Devanagari file)

Metric	Value
Median question length	6 words (min 3, max 12)
Median answer length	5 words (min 1, max 24)
Questions containing Latin text (code-mixing)	98 of 262
Questions containing digits	35 of 262
Answers containing digits	105 of 262 (40.1%)

Question words (approximate substring counts; one question can match several):

Question word	Approximate count	Meaning
को	148	who / whose
के	111	what
कति	74	how much / how many
किन	24	why
कहाँ	17	where
कुन	13	which
कसरी	12	how
कहिले	11	when
कसले	3	by whom

5.4 Example pairs

Romanized question	Romanized answer
sahas urja limited kaha awasthit chha?	hattisar, kathamandau-1, nepalma awasthit chha.
sahas urja limitedko samparka phone number ke ho?	01-4419182, 4419183 ra 4422756.
prati sheyar ankit mulya kati ho?	ru. 100/-.
laganima hune jokhimko mulyankana kasko jimmewari ho?	laganikarta swayanko pramukh dayitwa ho.

The Devanagari file has the same ids and the same facts, so the Devanagari version of each pair is found by matching id.

---

6. Dataset 3: Mountain Energy Nepal Annual Report FY 2081/82

6.1 What it contains

2,000 QA pairs from the 12th Annual Report of Mountain Energy Nepal Ltd. for fiscal year 2081/82. Topics include the board and governance, the chairman's report, audit and financial statements, projects, shareholding, dividend and bonus share proposals, the general meeting, and the registered office.

6.2 Key design feature: one fact asked in many styles

The 2,000 questions are answered by only 360 unique answers. Of these, 356 answers repeat and together cover 1,996 of the 2,000 rows. The same fact is asked in several phrasings, so the dataset tests robustness to wording rather than breadth of facts.

Example: the registered office fact has 17 rows, all with the same answer.

Romanized question	Style
kampaniko rajishtarda karyalaya kaha chha?	plain
sankshepma bhannuhos, kampaniko rajishtarda karyalaya kaha chha?	brief
spashta parnuhos: kampaniko rajishtarda karyalaya kaha chha?	explain
sanchalak samitiko pratiwedanama kampaniko thegana purna rupma ke ho?	report-anchored

All of them have the same answer: ka.ma.na.pa. wada nan. 30, dillibajar, kathamadaun.

Styles can also be stacked. In the romanized Men file, 257 questions combine two or more style markers, for example ma janna chahanchhu, sankshepma bhannuhos, ... or ke tapai bhanna saknuhunchha sankshepma bhannuhos, ... spashta parnuhos?. The opener-based counts in Section 6.3 count only the leading style, so these stacked rows are not reflected in those numbers. Across the whole file, 1,137 questions contain at least one style marker anywhere.

6.3 Question style variants

Styles are identified by opening phrase. The counts below are identical in the Devanagari and romanized files.

Style	Devanagari opener	Romanized opener	Count	Avg question words	Avg answer words
Plain direct question (incl. one-sentence style below)	(none)	(none)	1,161	11.2	5.9
One sentence	एक वाक्यमा भन्नुहोस्,	eka wakyama bhannuhos,	62	13.0	6.3
Report-anchored	प्रतिवेदन अनुसार ...	pratiwedana anusar ...	103	11.5	5.7
Imperative	जानकारी दिनुहोस्:	janakari dinuhos:	103	10.6	5.8
Brief request	संक्षेपमा भन्नुहोस्,	sankshepma bhannuhos,	103	10.7	6.0
Yes/no framing	के तपाईं भन्न सक्नुहुन्छ ...?	ke tapai bhanna saknuhunchha ...?	103	13.3	6.3
Explain	स्पष्ट पार्नुहोस्:	spashta parnuhos:	102	10.8	5.6
Polite request	कृपया बताउनुहोस्,	kripaya bataunuhos,	101	10.7	6.0
First person	म जान्न चाहन्छु ...	ma janna chahanchhu ...	100	11.5	5.5
No detail	विस्तारमा नभई मुख्य तथ्य मात्र:	wistarma nabhai mukhya tathya matra:	62	16.0	6.0

The "Avg" columns come from the romanized file. The one-sentence row shows the Devanagari averages, and the plain row includes the one-sentence rows in its average (11.2 and 5.9).

Answer length is almost the same across styles (5.5 to 6.3 words). The style changes the question, not the expected answer.

6.4 Content statistics

Metric	Value
Median question length	11 words (min 3, max 23)
Median answer length	5 words (min 1, max 18)
Answers containing digits	1,430 of 2,000 (71.5%)
Answers containing rupee amounts (रु. / ru.)	685 (Devanagari count)
Answers referencing a page number	82 (Devanagari count)
Questions referencing a fiscal year	300 (Devanagari count)

Most repeated answers:

Times	Answer (romanized)
17	ka.ma.na.pa. wada nan. 30, dillibajar, kathamadaun.
15	65 me.wa. ho.
9	sam bahadur bhatta kampani sachiv hun.
9	rajendra prasad sharma hun.
9	da. harishankar shreshthale miti 2081/10/22 ma rajinama die.

6.5 Example pairs

Romanized question	Romanized answer
mauntena inarji nepal limitedko adhyaksha ko hun?	shri pradip janga pande adhyaksha hun.
mauntena inarji nepalka sanchalak samiti sadasyaharu ko-ko hun?	adhyaksha pradip janga pande, sanchalak arun kumar ke.si., tika prasad upreti, jhanak datta khanal ra rachana adhikari hun.
kampaniko phone number ke ho?	01-4537705, 01-4541648 ho.
aa.wa. 2081/82 ma prashasanik kharcha kati thiyo?	ru. 47,544,765 thiyo.

---

7. Diversity Analysis

7.1 Method and a correction

All diversity numbers in this README use one consistent tokenizer: lowercase the text, remove punctuation, and split on whitespace. Earlier versions of this README used a regex tokenizer (\w+) that splits Devanagari words into fragments, because Devanagari vowel signs are not word characters in that regex. That made the Devanagari metrics wrong. The corrected values are below. The romanized metrics were not affected in the same way.

Type-token ratio (TTR) is distinct tokens divided by total tokens. Distinct bigram ratio is distinct two-word sequences divided by total two-word sequences. Higher means more variety. Ratios depend on file size, so compare files of similar size with care.

7.2 Lexical diversity (corrected)

File	Rows	Unique Q	Unique A	Question vocab	Answer vocab	TTR (Q)	TTR (A)	Bigram ratio (Q)	Median Q words	Median A words
Bank (romanized)	25,472	25,464	25,469	2,006	8,763	0.0086	0.0360	0.0324	9	9
Sahas (Devanagari)	262	262	256	497	723	0.2792	0.4516	0.6331	6	5
Sahas (romanized)	262	262	256	497	733	0.2792	0.4573	0.6331	6	5
Mountain Energy (Devanagari)	2,000	2,000	360	671	1,032	0.0284	0.0710	0.0893	11	5
Mountain Energy (romanized)	2,000	2,000	360	669	1,027	0.0283	0.0706	0.0893	11	5

Observations from the corrected numbers:

• The Devanagari and romanized versions of Sahas and Mountain Energy have essentially the same lexical diversity. The romanization preserves the structure of the text.
• Sahas is the most varied per token of the three datasets, because every question is written individually.
• Mountain Energy has low per-token variety because the same facts are asked many times.
• The bank dataset has the lowest question variety, because of its 8 fixed templates.

7.3 Behaviour diversity

Level	Bank	Sahas	Mountain Energy	Overall
Distinct metadata behavior labels	1	1	1	2 labels in total (English and Nepali, both meaning "factual answer")
Distinct question_type labels	1	1	1	1 (fact-based)
Question style variants	8 fixed templates (one per column)	Free-form	10 styles (Section 6.3)	Varies by file
Answer length variation	Narrow (about 9 words)	Wide (1 to 24 words)	Narrow (5.5 to 6.3 words by style)	Low

The metadata captures almost none of the behavioural variation. All files are labelled as factual QA. The real variation is in how questions are phrased (Mountain Energy) and which column is asked (bank). For behavioural analysis, use the question text rather than the behavior field.

7.4 Question pattern diversity

Metric	Bank	Sahas	Mountain Energy
Fixed question templates	8	None (hand-written)	Style prefixes (10), with the same facts repeated
Question word openers	Bank names, then a template tail	Varied wh-words (Section 5.3)	Style openers plus fiscal-year or company phrases
Yes/no framing	None	None	103 rows (ke tapai ...) answered with a fact, not yes/no

7.5 Answer diversity

Metric	Bank	Sahas	Mountain Energy
Unique answers	25,469	256	360
Answers containing digits	37.0%	40.1%	71.5%
Answer types	Dates, numbers, categories, places	Numbers, names, short phrases, rupee amounts	Rupee amounts, names, counts, addresses, dates

Mountain Energy has very few unique answers (360 for 2,000 rows). This is intentional: it is a robustness set, not a broad fact set.

7.6 Bank answer types by column group

Answer type	Columns	Count
Date	Failure Date	3,184
Number	Total Deposits, Total Assets	6,368
Category or text	Headquarters, Institution Type, Charter Type, Insurance Fund, Transaction Type	15,920

7.7 Overall rating (qualitative)

Aspect	Bank	Sahas	Mountain Energy
Behavioural diversity	Low	Low	Medium (style variants only)
Question pattern diversity	Low	High	Medium
Lexical diversity	Low	High	Low
Fact diversity	High (one fact per bank and column)	Medium	Low
Answer diversity	Medium	High	Low
Script diversity	Romanized only	Both forms	Both forms

Ratings are qualitative and based on the metrics above.

---

8. Known Issues and Limitations

Please read this section before using the data in production.

#	Dataset	Issue	Impact	Suggested handling
1	Bank	bank_failures_ROMANIZED_qa_1.jsonl is byte-identical to bank_failures_ROMANIZED_qa.jsonl (same MD5).	Loading both doubles the bank data and leaks everything across any split.	Load only one file.
2	All romanized files	The metadata fields script (Deva), content_script (देवनागरी), content_language (नेपाली), and behavior_definition (converting to "pure Nepali") describe a Devanagari target, but the romanized text is Latin.	Filtering by script or content_script returns the wrong set for the romanized files.	Do not rely on these fields to identify script. Detect script from the text, or use the file name.
3	Mountain Energy (romanized)	39 questions (and 4 of their answers) contain two words that are mostly Latin but still include one Devanagari character, left over from romanization. praॉphit (28 occurrences, likely profit) and praॉparti (15 occurrences, likely property). The stray character is the Devanagari candra-o sign ॉ.	Tokens with these words are out of vocabulary for a Latin-only model.	Use the cleaning code in Section 9.8, or re-romanize. The uploaded files have not been changed. Confirm the intended words against the Devanagari file before replacing.
4	Bank	Dollar amounts for Total Deposits and Total Assets have no unit (for example 91, 1064).	Models may learn values without scale.	State the unit in the prompt, or check the FDIC source.
5	Bank	Dates use US format (M/D/YYYY), for example 4/19/1934.	Date comparisons break if parsed as day-first.	Parse with %m/%d/%Y.
6	Bank	3 exact duplicate QA pairs. 5 more questions appear twice with different answers, probably because two failed banks share a name.	Small leakage between splits if random.	Deduplicate exact pairs. Check the source for the 5 collisions.
7	Bank	Only 8 fixed question templates.	A model may learn the template rather than the facts.	Do not use this file alone to evaluate question understanding.
8	Mountain Energy (both forms)	Only 360 unique answers across 2,000 questions. The same fact appears up to 17 times.	A random split puts the same answer in train and test, which inflates test scores.	Split by answer (Section 9.7).
9	Mountain Energy (both forms)	Yes/no framing (ke tapai ..., के तपाईं ...) is answered with the fact, not yes or no.	A model may learn that yes/no questions get factual answers.	Keep in mind when evaluating, or filter those rows.
10	Sahas (both forms)	Only 262 rows.	Too small to train from scratch.	Use for evaluation, few-shot examples, or as an addition to larger data.
11	Sahas and Mountain Energy (Devanagari)	U+200D (zero-width joiner) appears 3 times in Sahas and 50 times across 26 rows in Mountain Energy.	Valid Devanagari conjunct joiner, not an error.	Keep it. Strip only if you are deliberately normalizing.
12	Sahas and Mountain Energy (romanized)	Romanization is a custom scheme, not a standard transliteration. Proper nouns and terms are spelled by sound.	Spelling of names may not match official or English spellings.	Do not use romanized names as lookup keys without normalization.
13	All	All answers are short and factual.	A model trained on this alone may give short answers to open questions.	Mix with longer-form data for general chat use.
14	All	The license field says Apache-2.0, but the underlying documents are a public database, a company prospectus, and a company annual report.	The QA pair license does not automatically grant rights to the source text.	Check the source rights before redistribution (Section 11).
15	All	Not every answer has been checked against its source.	Some errors may exist.	Spot-check a random sample against the source before production use.

---

9. How to Load the Data

All examples assume the files are in DATA_DIR. They were tested against the uploaded files.

9.1 Plain Python

import json

DATA_DIR = "/path/to/data"

def read_jsonl(path):
    with open(path, encoding="utf-8") as fh:
        return [json.loads(line) for line in fh if line.strip()]

bank = read_jsonl(f"{DATA_DIR}/bank_failures_ROMANIZED_qa.jsonl")
sahas_rom = read_jsonl(f"{DATA_DIR}/sahas_urja_ipo_qa_final_unique_romanized.jsonl")
sahas_dev = read_jsonl(f"{DATA_DIR}/sahas_urja_ipo_qa_final_unique.jsonl")
men_rom = read_jsonl(f"{DATA_DIR}/men_ar_2081_82_sharegpt_2000_ROMANIZED.jsonl")
men_dev = read_jsonl(f"{DATA_DIR}/men_ar_2081_82_sharegpt_2000.jsonl")

print(len(bank), len(sahas_rom), len(men_rom))   # 25472 262 2000

row = bank[0]
print(row["conversations"][0]["value"])            # question
print(row["conversations"][1]["value"])            # answer
meta = json.loads(row["metadata_json"])            # metadata is a JSON string
print(meta["column_focus"])                        # Failure Date

9.2 Using the included loader

load_datasets.py (included with this README) loads either script variant and flattens each row to instruction and output.

from load_datasets import load_all, summarize

records = load_all("/path/to/data", script="romanized")
summarize(records)

print(records[0])

Choose the script with script="romanized" or script="devanagari". Bank has no Devanagari variant, so it is skipped with a warning in Devanagari mode.

Command line:

python load_datasets.py --data-dir /path/to/data --script romanized
python load_datasets.py --data-dir /path/to/data --script devanagari

Expected output (romanized):

Total records: 27734
Per dataset:
   25472  fdic_bank_failures
     262  sahas_urja_ipo_prospectus
    2000  Mountain Energy Nepal Ltd. 12th Annual Report FY 2081/82
Per behavior label (metadata):
   27472  Provide factual answer derived only from the dataset
     262  जानकारीमूलक उत्तर दिने

Expected output (Devanagari):

[warn] bank_failures has no devanagari variant; skipped
Total records: 2262

9.3 HuggingFace datasets

from datasets import load_dataset

DATA_DIR = "/path/to/data"

ds = load_dataset(
    "json",
    data_files={
        "bank": f"{DATA_DIR}/bank_failures_ROMANIZED_qa.jsonl",
        "sahas": f"{DATA_DIR}/sahas_urja_ipo_qa_final_unique_romanized.jsonl",
        "men": f"{DATA_DIR}/men_ar_2081_82_sharegpt_2000_ROMANIZED.jsonl",
    },
    split=None,
)

print(ds)
print(ds["bank"][0]["conversations"])

9.4 Pandas

import json
import pandas as pd

df = pd.read_json("/path/to/data/sahas_urja_ipo_qa_final_unique_romanized.jsonl", lines=True)

df["question"] = df["conversations"].apply(lambda c: c[0]["value"])
df["answer"] = df["conversations"].apply(lambda c: c[1]["value"])
df["question_type"] = df["metadata_json"].apply(lambda s: json.loads(s)["question_type"])

print(df[["question", "answer"]].head())

9.5 Convert to instruction or chat format

def to_alpaca(row):
    return {
        "instruction": row["conversations"][0]["value"],
        "input": "",
        "output": row["conversations"][1]["value"],
    }

def to_chat(row):
    return {
        "messages": [
            {"role": "user", "content": row["conversations"][0]["value"]},
            {"role": "assistant", "content": row["conversations"][1]["value"]},
        ]
    }

alpaca_bank = ds["bank"].map(to_alpaca, remove_columns=ds["bank"].column_names)
chat_men = ds["men"].map(to_chat, remove_columns=ds["men"].column_names)

9.6 Filter by style or by column

import json

STYLE_PREFIXES = {
    "brief": "sankshepma",
    "polite": "kripaya",
    "imperative": "janakari dinuhos:",
    "explain": "spashta parnuhos:",
    "first_person": "ma janna",
    "yes_no": "ke tapai",
    "one_sentence": "eka wakyama",
    "no_detail": "wistarma nabhai",
    "report_anchored": "pratiwedana anusar",
}

def style_of(question):
    q = question.lower()
    for name, prefix in STYLE_PREFIXES.items():
        if q.startswith(prefix):
            return name
    return "plain"

men = ds["men"].add_column("style", [style_of(r["conversations"][0]["value"]) for r in ds["men"]])
print(men.filter(lambda r: r["style"] == "brief").num_rows)    # 103

bank_dates = ds["bank"].filter(
    lambda r: json.loads(r["metadata_json"])["column_focus"] == "Failure Date"
)
print(bank_dates.num_rows)    # 3184

9.7 Leak-safe train/test split (Mountain Energy)

Because one answer appears under many questions, a random split leaks facts. This split keeps all rows that share an answer in the same split.

import random
from datasets import load_dataset

men = load_dataset("json", data_files={
    "men": "/path/to/data/men_ar_2081_82_sharegpt_2000_ROMANIZED.jsonl",
}, split="men")

fact_id = {}
groups = []
for row in men:
    answer = row["conversations"][1]["value"]
    fact_id.setdefault(answer, len(fact_id))
    groups.append(fact_id[answer])

rng = random.Random(42)
unique_groups = list(set(groups))
rng.shuffle(unique_groups)
test_groups = set(unique_groups[: int(0.1 * len(unique_groups))])

is_test = [g in test_groups for g in groups]
train = men.select([i for i, t in enumerate(is_test) if not t])
test = men.select([i for i, t in enumerate(is_test) if t])

print(len(train), len(test))   # 1791 209

train_answers = {r["conversations"][1]["value"] for r in train}
test_answers = {r["conversations"][1]["value"] for r in test}
print(len(train_answers & test_answers))   # 0

9.8 Detect and fix the residual Devanagari (Mountain Energy romanized)

This finds rows that still contain Devanagari, and replaces the two known words. The replacements are the likely intended words (profit, property). Confirm them against the Devanagari file before using them in a final dataset. This code does not modify the uploaded files.

import re
from datasets import load_dataset

men = load_dataset("json", data_files={
    "men": "/path/to/data/men_ar_2081_82_sharegpt_2000_ROMANIZED.jsonl",
}, split="men")

DEVANAGARI = re.compile(r"[\u0900-\u097F]")

def has_residual(row):
    return any(DEVANAGARI.search(t["value"]) for t in row["conversations"])

print(len(men.filter(has_residual)))   # 39

REPLACEMENTS = {"praॉphit": "profit", "praॉparti": "property"}

def fix_residual(row):
    for turn in row["conversations"]:
        for wrong, right in REPLACEMENTS.items():
            turn["value"] = turn["value"].replace(wrong, right)
    return row

men_clean = men.map(fix_residual)
print(len(men_clean.filter(has_residual)))   # 0

9.9 Verify that the two forms are aligned

from datasets import load_dataset

dev = load_dataset("json", data_files={"s": "/path/to/data/sahas_urja_ipo_qa_final_unique.jsonl"}, split="s")
rom = load_dataset("json", data_files={"s": "/path/to/data/sahas_urja_ipo_qa_final_unique_romanized.jsonl"}, split="s")

print(dev["id"] == rom["id"])    # True

9.10 Remove exact duplicates (bank)

from datasets import load_dataset

bank = load_dataset("json", data_files={
    "bank": "/path/to/data/bank_failures_ROMANIZED_qa.jsonl",
}, split="bank")

seen = set()
keep = []
for i, row in enumerate(bank):
    key = (row["conversations"][0]["value"], row["conversations"][1]["value"])
    if key not in seen:
        seen.add(key)
        keep.append(i)

bank_unique = bank.select(keep)
print(len(bank_unique))   # 25469

---

10. Recommended Usage

Goal	Recommended file	Notes
Nepali factual recall at scale	Bank (romanized)	Repetitive wording. Mix with other data.
Robustness to phrasing	Mountain Energy (either form)	Use the split in Section 9.7.
Small evaluation set on corporate documents	Sahas (either form)	Only 262 rows, so report confidence intervals.
Comparing scripts	Sahas or Mountain Energy, both forms	Same ids, so the only difference is the script.
Romanized-input models	Bank, Sahas romanized, Mountain Energy romanized	Clean the residual first (Section 9.8).
Yes/no questions	Do not use the Mountain Energy yes/no rows as-is	Answers are factual, not yes or no.

General advice:

• Do not evaluate on facts you trained on, especially in Mountain Energy.
• Report results per dataset, since difficulty and style differ.
• Pick one script per training run unless you are studying cross-script transfer.

---

11. Licensing and Sources

Dataset	Original source	Publisher	License field	Points to check
Bank Failures	https://www.kaggle.com/datasets/fdic/bank-failures	FDIC (US Federal Deposit Insurance Corporation)	Apache-2.0	Confirm the terms on the Kaggle page.
Sahas Urja IPO	https://sahasurja.com/wp-content/uploads/2021/08/Prospectus-for-IPO.pdf	Sahas Urja Limited	Apache-2.0	Company prospectus. Check reuse terms.
Mountain Energy AR	https://mountainenergy.com.np/reportsAnnual.html (in the source field; url is empty)	Mountain Energy Nepal Ltd.	Apache-2.0	Company annual report. Check reuse terms.

The license field describes the QA pairs as provided. It does not change the rights of the underlying documents. Confirm those rights before redistributing the data or publishing a model trained on it.

---

12. Reproducing the Statistics

This script uses the same tokenizer as Section 7.1.

import json, re, statistics

PUNCT = re.compile(r"[?.,।!:;()'\"/\-–]+")

def tokens(text):
    return [t for t in PUNCT.sub(" ", text.lower()).split() if t]

def load(path):
    return [json.loads(l) for l in open(path, encoding="utf-8") if l.strip()]

for name, path in [
    ("bank", "bank_failures_ROMANIZED_qa.jsonl"),
    ("sahas_dev", "sahas_urja_ipo_qa_final_unique.jsonl"),
    ("sahas_rom", "sahas_urja_ipo_qa_final_unique_romanized.jsonl"),
    ("men_dev", "men_ar_2081_82_sharegpt_2000.jsonl"),
    ("men_rom", "men_ar_2081_82_sharegpt_2000_ROMANIZED.jsonl"),
]:
    rows = load(path)
    Q = [r["conversations"][0]["value"] for r in rows]
    A = [r["conversations"][1]["value"] for r in rows]
    qt = [t for q in Q for t in tokens(q)]
    at = [t for a in A for t in tokens(a)]
    bg = [tuple(tokens(q)[i:i+2]) for q in Q for i in range(len(tokens(q)) - 1)]
    print(f"{name:10s} rows={len(rows)} uQ={len(set(Q))} uA={len(set(A))} "
          f"TTR_q={len(set(qt))/len(qt):.4f} TTR_a={len(set(at))/len(at):.4f} "
          f"bigram_q={len(set(bg))/len(bg):.4f} "
          f"medQ={statistics.median(len(q.split()) for q in Q)} "
          f"medA={statistics.median(len(a.split()) for a in A)}")

Expected output matches Section 7.2 (values round to four decimals).

---

13. Changelog

• v2 (current)
  • Added the romanized versions of Sahas Urja and Mountain Energy. Verified row alignment, digit preservation, and id matching against the Devanagari files.
  • Found that bank_failures_ROMANIZED_qa_1.jsonl is an exact duplicate of the original bank file.
  • Found 39 Mountain Energy romanized rows with leftover Devanagari characters (praॉphit, praॉparti). Added detection and fix code (Section 9.8). The uploaded files are unchanged.
  • Corrected the Devanagari diversity metrics in v1. The earlier regex tokenizer split Devanagari words into fragments, which made the Sahas and Mountain Energy TTR and bigram values wrong. Corrected values are in Section 7.2.
  • Updated the loader to choose the script variant.
• v1: Initial release. Devanagari Sahas and Mountain Energy, romanized bank.

ShareSansar Romanized Nepali Stock-Market QA (30.5k)

A single-turn, question-answer dataset of 30,500 factual stock-market questions and answers about daily trading data from the Nepal Stock Exchange (NEPSE), as scraped by ShareSansar. Questions and answers are written in Romanized Nepali (Nepali written in Latin letters, for example banda mulya, kati thiyo?) and stored in the ShareGPT-style conversations format, one JSONL record per line.

Every answer is a short, complete sentence that restates the question and gives exactly one number taken from the source table. There is no reasoning, no advice and no free-form text, which makes the dataset good for factual look-up, numeric-answer formatting and Romanized Nepali domain adaptation.

Quick Facts

Property	Value
File	sharesansar_romanized_qa_30k_1.jsonl
File size	47,725,172 bytes (about 45.5 MiB)
Records	30,500 (one per line, UTF-8)
Turns per record	2 (one human, one gpt)
Format	JSON Lines, ShareGPT-style conversations
Language of text	Romanized Nepali (Latin script, ASCII only)
Domain	Stock market, Nepal Stock Exchange daily trading data
Task	Instruction following, short factual question answering
Distinct tickers (symbols)	546
Distinct trading dates	1,690 (2021-09-01 to 2026-07-19)
Distinct question targets (column_focus)	22
Distinct question templates	21
Distinct behaviours	1 (see Behaviour Diversity)
Generation type	real (values come from real scraped data, templated into text)
License tag	Apache-2.0 (inherited from the source repository, see License)

Contents

#	Section	What it covers
1	Source and Provenance	Where the numbers come from and how the text was built
2	Data Format and Schema	Every field, with types and values
3	Behaviour Diversity	Behaviour label, definition and what actually varies
4	Question Pattern Diversity	All 21 question templates, grouped, with glossary
5	Answer Pattern Diversity	Answer templates and number formats
6	Content Diversity	Tickers, dates, windows, value ranges
7	Statistics	Lengths, per-year counts, weekday spread
8	Quality Checks	Duplicates, ids, ordering, character set
9	How to Load the Dataset	Code demos: pandas, Hugging Face, plain Python, fine-tuning prep
10	Example Records	One real record per question target
11	Intended Uses	Good fits and poor fits
12	Known Limitations	Things to know before training
13	Suggested Splits	Leakage-safe splitting
14	License and Citation	Source license and citation stub

---

1. Source and Provenance

Item	Value
Source description	ShareSansar Nepal Stock Exchange Daily Data (GitHub Scrape)
Source name (source_name)	sharesansar_all_data
Source repository (source_repo)	sbmagar13/sharesansar_datascrape
Source URL (url)	https://github.com/sbmagar13/sharesansar_datascrape/tree/main/.github
Source config / split / revision	default / train / v1
Source row reference	source_row_id (identical to id in every record)
Position in the original table	original_row_index inside metadata_json (30,500 distinct values, range 0 to 487,984)
Generation type	real (every number is a real value from the scraped table)
Condition	real
How text is produced	Each record fills one fixed question template and one fixed answer template with a ticker, a date, and one cell value from the source row
Sampling from the source	The source table has far more rows than this file; original_row_index runs up to 487,984, and each record points to exactly one source row and one column

---

2. Data Format and Schema

Each line of the file is one JSON object with 19 top-level fields. Every record has exactly the same keys.

2.1 Top-level fields

Field	Type	Distinct values	Value (or example)	Meaning
id	string	30,500 (unique)	sharesansar_faq_00000	Unique record id, zero-padded counter
conversations	list of 2 objects	n/a	see 2.2	The question and the answer
source	string	1	ShareSansar Nepal Stock Exchange Daily Data (GitHub Scrape)	Human-readable source name
source_name	string	1	sharesansar_all_data	Short source key
source_repo	string	1	sbmagar13/sharesansar_datascrape	Upstream GitHub repository
source_config	string	1	default	Upstream config name
source_split	string	1	train	Upstream split name
source_revision	string	1	v1	Upstream revision
source_row_id	string	30,500	same as id	Row reference in the source collection
language	string	1	ne	Declared language (Nepali)
language_code	string	1	npi	Declared language code (Nepali, individual language)
script	string	1	Deva	Declared script (Devanagari), but the text is Latin, see Section 12
license	string	1	Apache-2.0	License tag
license_tier	string	1	permissive	License category
task_type	string	1	instruction-following	Task label
generation_type	string	1	real	Origin of the facts
condition	string	1	real	Condition label
url	string	1	GitHub URL of the source	Source location
metadata_json	string (JSON-encoded object)	n/a	see 2.3	Extra metadata, stored as a string and must be parsed with json.loads

2.2 The conversations field

Position	Key from	Key value	Notes
0	human	The question (Romanized Nepali, 43 to 78 characters)	Always a single sentence ending in ?
1	gpt	The answer (Romanized Nepali, 41 to 70 characters)	Always a single sentence ending in .

All 30,500 records have exactly this human then gpt order. There is no system prompt, no multi-turn dialogue and no context passage.

2.3 The metadata_json field (after parsing)

Key	Distinct values	Value / meaning
generation_domain	1	Stock Market
generation_category	1	Nepal Stock Exchange Daily Trading Data
generation_sub_domain	1	ShareSansar Price & Volume Reports
behavior	1	Provide factual answer derived only from the dataset
behavior_definition	1	Nepali (Devanagari) text, translated in Section 3
question_type	1	fact-based
content_language	1	नेपाली (Nepali)
content_script	1	देवनागरी (Devanagari)
english_content_allowed	1	false
column_focus	22	The source column the question asks about, the only metadata key that varies
original_row_index	30,500	Row position in the source table

---

3. Behaviour Diversity

3.1 The behaviour label

Property	Value
Behaviour (behavior)	Provide factual answer derived only from the dataset
Behaviour definition (behavior_definition), original	मूल प्रश्नको अर्थ र मूल उत्तरको तथ्य कायम राख्दै शुद्ध नेपालीमा रूपान्तरण गरिएको उत्तर दिने।
Behaviour definition, English translation	Give an answer converted into pure Nepali that keeps the meaning of the original question and the facts of the original answer.
Question type (question_type)	fact-based
Records with this behaviour	30,500 of 30,500 (100%)

3.2 What kinds of behaviour does the dataset contain?

The dataset is intentionally single-behaviour: every record is a closed-book-style fact look-up that returns one number. The behaviour label itself does not vary. The diversity lies in what is asked about, which is captured by column_focus and the question templates.

Diversity dimension	Distinct values	Varies?	Where to see it
Behaviour (behavior)	1	No	Section 3.1
Question type (question_type)	1 (fact-based)	No	Section 3.1
Task type (task_type)	1 (instruction-following)	No	Section 2.1
Turn structure	1 (single turn, human then gpt)	No	Section 2.2
Answer style	1 (restating sentence with one number)	No	Section 5
Question target (column_focus)	22	Yes	Section 4.2
Question template	21	Yes	Section 4.2
Answer template	21	Yes	Section 5.1
Ticker symbol	546	Yes	Section 6.1
Trading date	1,690	Yes	Section 6.2
Look-back window (52 / 120 / 180)	3	Yes	Section 6.3
Value kind (price, count, percent, signed difference, index)	5	Yes	Section 6.4

3.3 Reasoning skills exercised by the single behaviour

Because the behaviour is a fact look-up, the model is exercised on the following skills rather than on different behaviours.

Skill	How it appears in the data
Entity recognition	The ticker symbol at the start of the question must be copied into the answer
Date handling	An ISO date (YYYY-MM-DD) must be copied into the answer unchanged
Metric disambiguation	22 similar Romanized terms must be told apart (for example banda mulya vs aghillo banda mulya)
Numeric formatting	Thousands separators, two-decimal values, negative numbers and % signs must be reproduced
Verb agreement	kati thiyo? is answered with thiyo; kati pugyo? is answered with pugeko thiyo
Window numbers	The 52, 120 or 180 in the question must reappear in the answer

---

4. Question Pattern Diversity

4.1 Anatomy of a question

Every question follows one skeleton:

Slot	Content	Example
1. Ticker	Lowercase symbol (one record family uses uppercase API)	aclbsl
2. Linker	The fixed word ko ("of")	ko
3. Date	ISO date	2021-09-01
4. Linker	The fixed word ma ("on / in")	ma
5. Metric phrase	One of 21 fixed Romanized phrases (some with an English abbreviation in parentheses)	banda mulya (close)
6. Question tail	kati thiyo? (20 templates) or kati pugyo? (1 template)	kati thiyo?

Skeleton: <ticker> ko <date> ma <metric phrase> kati thiyo?

4.2 All 21 question templates (with counts)

This table lists every question template in the dataset. <SYM> is the ticker, <DATE> is the date, <N> is the window length. Templates map one-to-one to column_focus values, except that 120 Days and 180 Days share one template, so there are 22 targets but 21 templates.

#	Question template	Target(s) (column_focus)	Records
1	<SYM> ko <DATE> ma banda mulya (close) kati thiyo?	Close	4,206
2	<SYM> ko <DATE> ma <N> dinko ausat mulya kati thiyo?	120 Days, 180 Days	3,274
3	<SYM> ko <DATE> ma <N> haptako uchcha mulya kati thiyo?	52 Weeks High	1,471
4	<SYM> ko <DATE> ma khulne mulya (open) kati thiyo?	Open	1,387
5	<SYM> ko <DATE> ma uchchatam mulya (high) kati pugyo?	High	1,387
6	<SYM> ko <DATE> ma nyunatam mulya (low) kati thiyo?	Low	1,387
7	<SYM> ko <DATE> ma vwap mulya kati thiyo?	VWAP	1,387
8	<SYM> ko <DATE> ma karobar matra (volume) kati thiyo?	Vol	1,387
9	<SYM> ko <DATE> ma aghillo banda mulya (prev. close) kati thiyo?	Prev. Close	1,387
10	<SYM> ko <DATE> ma karobar rakam (turnover) kati thiyo?	Turnover	1,387
11	<SYM> ko <DATE> ma karobar sankhya (transactions) kati thiyo?	Trans.	1,386
12	<SYM> ko <DATE> ma mulya pharak (diff) kati thiyo?	Diff	1,386
13	<SYM> ko <DATE> ma mulya dayara (range) kati thiyo?	Range	1,386
14	<SYM> ko <DATE> ma mulya pariwartana pratishat (diff %) kati thiyo?	Diff %	1,386
15	<SYM> ko <DATE> ma mulya dayara pratishat (range %) kati thiyo?	Range %	1,386
16	<SYM> ko <DATE> ma vwap pratishat kati thiyo?	VWAP %	1,386
17	<SYM> ko <DATE> ma <N> haptako nyuna mulya kati thiyo?	52 Weeks Low	1,386
18	<SYM> ko <DATE> ma wishwas suchak (conf.) kati thiyo?	Conf.	799
19	<SYM> ko <DATE> ma antim karobar mulya (ltp) kati thiyo?	LTP	447
20	<SYM> ko <DATE> ma close ra ltp bichko pharak kati thiyo?	Close - LTP	446
21	<SYM> ko <DATE> ma close-ltp pratishat pharak kati thiyo?	Close - LTP %	446

4.3 Question families (grouping of the 22 targets)

Family	Targets (column_focus)	Records	Share
Price levels	Open, High, Low, Close, Prev. Close, LTP	10,201	33.45%
Traded volume and value	Vol, Turnover, Trans.	4,160	13.64%
Weighted average price	VWAP	1,387	4.55%
Change, range and spread	Diff, Range, Diff %, Range %, VWAP %, Close - LTP, Close - LTP %	7,822	25.65%
Long look-back statistics	120 Days, 180 Days, 52 Weeks High, 52 Weeks Low	6,131	20.10%
Indicator	Conf.	799	2.62%

4.4 Surface-form variation inside the templates

Variation	Observation	Records affected
Question tail	kati thiyo?	29,113
Question tail	kati pugyo? (only for High)	1,387
English abbreviation in parentheses, for example (close), (ltp), (diff %)	Present	20,704
No parenthesised abbreviation (VWAP, window averages, 52-week values, Close-LTP)	Present	9,796
Number inside the question (52, 120 or 180)	Present	6,131
Ticker case	Lowercase, except API (uppercase)	101 records contain an uppercase character
Question length	8 to 11 words, 43 to 78 characters	all
Paraphrase variants of the same question	None, each target has exactly one wording	all

4.5 Romanized glossary of the question vocabulary

Use this table to understand what the Romanized words mean.

Romanized word or phrase	English meaning	Used in
ko	of	every question
ma	on / in	every question
kati thiyo?	how much was it?	20 templates
kati pugyo?	how much did it reach?	High
banda mulya	closing price	Close
khulne mulya	opening price	Open
uchchatam mulya	highest price	High
nyunatam mulya	lowest price	Low
aghillo banda mulya	previous closing price	Prev. Close
antim karobar mulya	last traded price	LTP
vwap mulya	volume-weighted average price	VWAP
karobar matra	traded quantity (volume)	Vol
karobar rakam	traded amount (turnover)	Turnover
karobar sankhya	number of transactions	Trans.
mulya pharak	price difference	Diff
mulya dayara	price range	Range
pratishat	percent	all % targets
mulya pariwartana pratishat	percent change in price	Diff %
mulya dayara pratishat	range as a percent	Range %
vwap pratishat	VWAP as a percent	VWAP %
wishwas suchak	confidence indicator (literally "trust index"), shown in the source as Conf.	Conf.
close ra ltp bichko pharak	difference between close and LTP	Close - LTP
close-ltp pratishat pharak	percent difference between close and LTP	Close - LTP %
<N> dinko ausat mulya	average price over the last N days	120 Days, 180 Days
<N> haptako uchcha mulya	highest price over the last N weeks	52 Weeks High
<N> haptako nyuna mulya	lowest price over the last N weeks	52 Weeks Low

4.6 Time-dependent availability of question targets

Not every target is present for the whole period. Some targets only appear later in the date range, so their records start later.

Target group	Earliest date in the data	Latest date in the data
Core price and volume targets (Open, High, Low, Close, Vol, Turnover, Trans., Prev. Close, VWAP, Diff, Range, and the three percent targets, Conf.)	2021-09-01	2026-07-19
Look-back targets (120 Days, 180 Days, 52 Weeks High, 52 Weeks Low)	2021-11-03	2026-07-19
LTP targets (LTP, Close - LTP, Close - LTP %)	2025-03-26	2026-07-19

---

5. Answer Pattern Diversity

5.1 All 21 answer templates

Each answer repeats the ticker, the date and the metric phrase from the question, then gives the value followed by a verb.

Answer template	Records
<SYM> ko <DATE> ma banda mulya <VALUE> thiyo.	4,206
<SYM> ko <DATE> ma <N> dinko ausat mulya <VALUE> thiyo.	3,274
<SYM> ko <DATE> ma <N> haptako uchcha mulya <VALUE> thiyo.	1,471
<SYM> ko <DATE> ma khulne mulya <VALUE> thiyo.	1,387
<SYM> ko <DATE> ma uchchatam mulya <VALUE> pugeko thiyo.	1,387
<SYM> ko <DATE> ma nyunatam mulya <VALUE> thiyo.	1,387
<SYM> ko <DATE> ma vwap mulya <VALUE> thiyo.	1,387
<SYM> ko <DATE> ma karobar matra <VALUE> thiyo.	1,387
<SYM> ko <DATE> ma aghillo banda mulya <VALUE> thiyo.	1,387
<SYM> ko <DATE> ma karobar rakam <VALUE> thiyo.	1,387
<SYM> ko <DATE> ma karobar sankhya <VALUE> thiyo.	1,386
<SYM> ko <DATE> ma mulya pharak <VALUE> thiyo.	1,386
<SYM> ko <DATE> ma mulya dayara <VALUE> thiyo.	1,386
<SYM> ko <DATE> ma mulya pariwartana pratishat <VALUE>% thiyo.	1,386
<SYM> ko <DATE> ma mulya dayara pratishat <VALUE>% thiyo.	1,386
<SYM> ko <DATE> ma vwap pratishat <VALUE>% thiyo.	1,386
<SYM> ko <DATE> ma <N> haptako nyuna mulya <VALUE> thiyo.	1,386
<SYM> ko <DATE> ma wishwas suchak <VALUE> thiyo.	799
<SYM> ko <DATE> ma antim karobar mulya <VALUE> thiyo.	447
<SYM> ko <DATE> ma close ra ltp bichko pharak <VALUE> thiyo.	446
<SYM> ko <DATE> ma close-ltp pratishat pharak <VALUE>% thiyo.	446

5.2 Answer structure

Property	Value
Sentences per answer	1
Numbers per answer	1 value (plus the date, plus the window length for 52/120/180 targets)
Ending	thiyo. (20 templates) or pugeko thiyo. (High only)
Percent sign	Attached to the value for Diff %, Range %, VWAP % and Close - LTP % (for example -0.78%)
Answer length	8 to 11 words, 41 to 70 characters
Extra explanation, units or currency	None
Refusals, "not available", or hedging	None

5.3 Number formatting rules found in the answers

Rule	Example	Notes
Thousands separator is a comma	1,581, 190,853	Used for values of 1,000 or more
Decimal separator is a dot	1,192.39	Two decimals where the source has decimals
Trailing zero is kept	3.50, 25,177.50	Trailing zeros are preserved
Integers have no decimals	742, 395	Volume and transactions are always integers
Negative numbers use a leading minus	-0.78%, -484.20	Appear in the signed targets only
Percent targets end in %	4.68%	Four targets

---

6. Content Diversity

6.1 Tickers (symbols)

Property	Value
Distinct tickers	546
Records per ticker, minimum / median / maximum	1 / 56.0 / 136
Tickers with fewer than 10 records	113
Records whose ticker contains a digit (for example adblb86, c30mf, bokd86ka)	3,568
Uppercase ticker	API only (101 records)

Ticker length distribution:

Ticker length (characters)	Records	Share
3	6,469	21.21%
4	12,400	40.66%
5	8,115	26.61%
6	2,526	8.28%
7	459	1.50%
8	285	0.93%
9	123	0.40%
10	108	0.35%
12	15	0.05%

Most frequent tickers:

Ticker	Records
sadbl	136
sicl	132
mhnl	131
ngpl	131
men	131
aclbsl	130
glbsl	128
cfcl	128
nil	127
vlbs	127

6.2 Dates

Property	Value
Date format	YYYY-MM-DD, always inside the question and the answer
Distinct dates	1,690
First date	2021-09-01
Last date	2026-07-19
Records per date, minimum / median / maximum	13 / 19.0 / 28
Ordering in the file	Sorted by date, ascending (no date goes backwards)

6.3 Look-back windows

Window length is part of the question text for four targets.

Target	Window number in question	Unit	Records
120 Days	120	days	1,936
180 Days	180	days	1,338
52 Weeks High	52	weeks	1,471
52 Weeks Low	52	weeks	1,386

6.4 Kinds of values asked for

Value kind	Targets	Typical characteristics
Price level	Open, High, Low, Close, Prev. Close, LTP, VWAP, 120 Days, 180 Days, 52 Weeks High, 52 Weeks Low	Positive decimals or integers, up to 50,500
Count or quantity	Vol, Trans.	Integers only, comma separated
Money amount	Turnover	Large decimals, up to about 551 million
Difference or span	Diff, Range, Close - LTP	Diff and Close - LTP can be negative, zero or positive; Range is zero or positive
Ratio, percent or index	Diff %, Range %, VWAP %, Close - LTP %, Conf.	Small decimals, percent sign for the four % targets

6.5 Value statistics per target

Computed from the numbers inside the answers.

Target	Records	Min	Median	Max	Has decimals	Negative	Zero
Open	1,387	7.15	487.50	49,500.00	40%	0	0
High	1,387	7.30	508.00	49,105.00	37%	0	0
Low	1,387	6.75	500.00	49,000.00	44%	0	0
Close	4,206	6.87	494.00	48,000.00	39%	0	0
Prev. Close	1,387	0.00	499.10	50,500.00	43%	0	2
LTP	447	8.40	527.00	46,888.00	46%	0	0
Vol	1,387	1.00	12,150.00	1,668,005.00	0%	0	0
Turnover	1,387	990.00	5,575,308.40	551,464,411.70	73%	0	0
Trans.	1,386	1.00	99.00	5,605.00	0%	0	0
VWAP	1,387	7.01	515.48	50,199.06	95%	0	0
Diff	1,386	-484.20	-0.14	1,475.60	58%	749	57
Range	1,386	0.00	14.90	981.20	56%	0	78
Diff %	1,386	-9.99	-0.18	14.98	93%	743	64
Range %	1,386	0.00	3.19	21.49	93%	0	78
VWAP %	1,386	-400.00	0.00	14.45	93%	656	90
Close - LTP	446	-57.62	0.00	51.17	28%	63	319
Close - LTP %	446	-0.02	0.00	0.10	7%	18	416
120 Days	1,936	0.00	453.77	46,660.93	90%	0	176
180 Days	1,338	0.00	452.50	47,101.29	88%	0	143
52 Weeks High	1,471	8.29	696.00	49,250.00	35%	0	0
52 Weeks Low	1,386	6.53	355.55	43,800.00	49%	0	0
Conf.	799	0.00	38.85	68.39	99%	0	1

---

7. Statistics

7.1 Length statistics

Measure	Min	Mean	Median	95th percentile	Max
Question characters	43	56.7	56	69	78
Answer characters	41	51.8	50	61	70
Question words	8	9.4	9	11	11
Answer words	8	8.7	8	10	11

All records are very short, so a full record fits easily in a small context window.

7.2 Records per year

Year	Records	Share	Note
2021	841	2.76%	Partial year, starts 2021-09-01
2022	5,053	16.57%	Full year
2023	5,972	19.58%	Full year
2024	7,186	23.56%	Full year
2025	7,230	23.70%	Full year
2026	4,218	13.83%	Partial year, ends 2026-07-19

7.3 Records per weekday

Weekday	Records	Share
Sunday	4,356	14.28%
Monday	4,364	14.31%
Tuesday	4,287	14.06%
Wednesday	4,400	14.43%
Thursday	4,387	14.38%
Friday	4,348	14.26%
Saturday	4,358	14.29%

7.4 Records per question target

Target (column_focus)	Family	Records	Share
Open	Price levels	1,387	4.55%
High	Price levels	1,387	4.55%
Low	Price levels	1,387	4.55%
Close	Price levels	4,206	13.79%
Prev. Close	Price levels	1,387	4.55%
LTP	Price levels	447	1.47%
Vol	Traded volume and value	1,387	4.55%
Turnover	Traded volume and value	1,387	4.55%
Trans.	Traded volume and value	1,386	4.54%
VWAP	Weighted average price	1,387	4.55%
Diff	Change, range and spread	1,386	4.54%
Range	Change, range and spread	1,386	4.54%
Diff %	Change, range and spread	1,386	4.54%
Range %	Change, range and spread	1,386	4.54%
VWAP %	Change, range and spread	1,386	4.54%
Close - LTP	Change, range and spread	446	1.46%
Close - LTP %	Change, range and spread	446	1.46%
120 Days	Long look-back statistics	1,936	6.35%
180 Days	Long look-back statistics	1,338	4.39%
52 Weeks High	Long look-back statistics	1,471	4.82%
52 Weeks Low	Long look-back statistics	1,386	4.54%
Conf.	Indicator	799	2.62%
Total		30,500	100.00%

---

8. Quality Checks

All checks below were run on the released file.

Check	Result
Total lines parsed as valid JSON	30,500 of 30,500
Identical key set in every record	Yes (19 top-level keys, 11 metadata keys)
Unique id values	30,500 (no duplicates)
id equals source_row_id	Yes, in every record
Unique original_row_index values	30,500
Duplicate questions	0
Duplicate question and answer pairs	0
Duplicate (ticker, date, target) combinations	0
Turn order is human then gpt	Yes, in every record
Answer value could be parsed with a single regular expression	30,500 of 30,500
Non-ASCII characters in questions or answers	0
Devanagari characters in questions or answers	0
Dates sorted ascending through the file	Yes

---

9. How to Load the Dataset

The file is plain JSON Lines, so it works with the standard library, pandas and Hugging Face datasets. Replace the path with your local path (or your Hub repo id).

9.1 Plain Python (no dependencies)

import json

PATH = "sharesansar_romanized_qa_30k_1.jsonl"

records = []
with open(PATH, encoding="utf-8") as f:
    for line in f:
        records.append(json.loads(line))

print(len(records))                      # 30500
r = records[0]
print(r["id"])                           # sharesansar_faq_00000
print(r["conversations"][0]["value"])    # question
print(r["conversations"][1]["value"])    # answer

# metadata_json is a STRING, parse it first
meta = json.loads(r["metadata_json"])
print(meta["column_focus"])              # Close
print(meta["behavior"])                  # Provide factual answer derived only from the dataset

9.2 pandas (flat table with parsed metadata)

import json
import pandas as pd

PATH = "sharesansar_romanized_qa_30k_1.jsonl"

df = pd.read_json(PATH, lines=True)

# 1) expand the JSON-encoded metadata into real columns
meta = pd.json_normalize(df["metadata_json"].apply(json.loads))
df = pd.concat([df.drop(columns=["metadata_json"]), meta], axis=1)

# 2) pull question and answer out of the conversation list
df["question"] = df["conversations"].str[0].str["value"]
df["answer"]   = df["conversations"].str[1].str["value"]

# 3) extract ticker, date and the numeric value
df["ticker"] = df["question"].str.split().str[0]
df["date"]   = pd.to_datetime(df["question"].str.extract(r"(\d{4}-\d{2}-\d{2})")[0])
df["value"]  = (
    df["answer"]
    .str.extract(r"(-?[\d,]+\.?\d*)%? (?:pugeko )?thiyo\.$")[0]
    .str.replace(",", "", regex=False)
    .astype(float)
)

print(df[["ticker", "date", "column_focus", "value"]].head())
print(df["column_focus"].value_counts())

9.3 Hugging Face datasets

from datasets import load_dataset

# local file
ds = load_dataset(
    "json",
    data_files="sharesansar_romanized_qa_30k_1.jsonl",
    split="train",
)
print(ds)
print(ds[0]["conversations"])

# from the Hub, after you upload the file (replace with your repo id)
# ds = load_dataset("your-username/sharesansar-romanized-qa-30k", split="train")

# metadata_json is a string; parse it inside a map() call
import json
ds = ds.map(lambda ex: {"column_focus": json.loads(ex["metadata_json"])["column_focus"]})

# filter by question target
close_only = ds.filter(lambda ex: ex["column_focus"] == "Close")
print(len(close_only))   # 4206

# streaming mode (no full download into memory)
stream = load_dataset(
    "json",
    data_files="sharesansar_romanized_qa_30k_1.jsonl",
    split="train",
    streaming=True,
)
print(next(iter(stream))["id"])

9.4 Filtering by target, ticker or date

import json

def load(path):
    with open(path, encoding="utf-8") as f:
        for line in f:
            r = json.loads(line)
            r["meta"] = json.loads(r["metadata_json"])
            r["q"] = r["conversations"][0]["value"]
            r["a"] = r["conversations"][1]["value"]
            yield r

rows = list(load("sharesansar_romanized_qa_30k_1.jsonl"))

# only volume questions
vol = [r for r in rows if r["meta"]["column_focus"] == "Vol"]

# only one ticker
nabil = [r for r in rows if r["q"].split()[0] == "nabil"]

# only 2025 questions (ISO dates compare correctly as strings)
y2025 = [r for r in rows if "2025-01-01" <= r["q"].split()[2] <= "2025-12-31"]

print(len(vol), len(nabil), len(y2025))

9.5 Convert to the messages chat format (OpenAI / Hugging Face style)

ROLE = {"human": "user", "gpt": "assistant"}

def to_messages(record):
    return {
        "id": record["id"],
        "messages": [
            {"role": ROLE[t["from"]], "content": t["value"]}
            for t in record["conversations"]
        ],
    }

# with Hugging Face datasets
from datasets import load_dataset
ds = load_dataset("json", data_files="sharesansar_romanized_qa_30k_1.jsonl", split="train")
chat_ds = ds.map(to_messages, remove_columns=ds.column_names)
print(chat_ds[0]["messages"])

To feed it to a tokenizer chat template:

from transformers import AutoTokenizer

tok = AutoTokenizer.from_pretrained("your-base-model")
text = tok.apply_chat_template(chat_ds[0]["messages"], tokenize=False)
print(text)

9.6 Date-based split (recommended)

import json

rows = [json.loads(l) for l in open("sharesansar_romanized_qa_30k_1.jsonl", encoding="utf-8")]
date_of = lambda r: r["conversations"][0]["value"].split()[2]

train = [r for r in rows if date_of(r) <  "2026-01-01"]
valid = [r for r in rows if "2026-01-01" <= date_of(r) < "2026-05-01"]
test  = [r for r in rows if date_of(r) >= "2026-05-01"]
print(len(train), len(valid), len(test))

9.7 Quick sanity-check report

import json, collections

rows = [json.loads(l) for l in open("sharesansar_romanized_qa_30k_1.jsonl", encoding="utf-8")]
targets = collections.Counter(json.loads(r["metadata_json"])["column_focus"] for r in rows)
for name, n in targets.most_common():
    print(f"{name:15} {n:6}")

---

10. Example Records

10.1 Full record as stored in the file

{
  "id": "sharesansar_faq_00000",
  "conversations": [
    {"from": "human", "value": "aclbsl ko 2021-09-01 ma banda mulya (close) kati thiyo?"},
    {"from": "gpt",   "value": "aclbsl ko 2021-09-01 ma banda mulya 1,581 thiyo."}
  ],
  "source": "ShareSansar Nepal Stock Exchange Daily Data (GitHub Scrape)",
  "source_name": "sharesansar_all_data",
  "source_repo": "sbmagar13/sharesansar_datascrape",
  "source_config": "default",
  "source_split": "train",
  "source_revision": "v1",
  "source_row_id": "sharesansar_faq_00000",
  "language": "ne",
  "language_code": "npi",
  "script": "Deva",
  "license": "Apache-2.0",
  "license_tier": "permissive",
  "task_type": "instruction-following",
  "generation_type": "real",
  "condition": "real",
  "url": "https://github.com/sbmagar13/sharesansar_datascrape/tree/main/.github",
  "metadata_json": "{\"generation_domain\": \"Stock Market\", \"generation_category\": \"Nepal Stock Exchange Daily Trading Data\", \"generation_sub_domain\": \"ShareSansar Price & Volume Reports\", \"behavior\": \"Provide factual answer derived only from the dataset\", \"behavior_definition\": \"(Nepali text, see Section 3.1)\", \"question_type\": \"fact-based\", \"content_language\": \"नेपाली\", \"content_script\": \"देवनागरी\", \"english_content_allowed\": false, \"column_focus\": \"Close\", \"original_row_index\": 0}"
}

(The behavior_definition value is shortened in this display only.)

10.2 One real question and answer for each of the 22 targets

Target	Question	Answer
Open	cbbl ko 2021-09-01 ma khulne mulya (open) kati thiyo?	cbbl ko 2021-09-01 ma khulne mulya 1,836 thiyo.
High	ebl ko 2021-09-01 ma uchchatam mulya (high) kati pugyo?	ebl ko 2021-09-01 ma uchchatam mulya 742 pugeko thiyo.
Low	glbsl ko 2021-09-01 ma nyunatam mulya (low) kati thiyo?	glbsl ko 2021-09-01 ma nyunatam mulya 1,680 thiyo.
Close	aclbsl ko 2021-09-01 ma banda mulya (close) kati thiyo?	aclbsl ko 2021-09-01 ma banda mulya 1,581 thiyo.
Prev. Close	mhnl ko 2021-09-01 ma aghillo banda mulya (prev. close) kati thiyo?	mhnl ko 2021-09-01 ma aghillo banda mulya 535 thiyo.
LTP	gmfbs ko 2025-03-26 ma antim karobar mulya (ltp) kati thiyo?	gmfbs ko 2025-03-26 ma antim karobar mulya 1,565.30 thiyo.
Vol	ksbbl ko 2021-09-01 ma karobar matra (volume) kati thiyo?	ksbbl ko 2021-09-01 ma karobar matra 190,853 thiyo.
Turnover	nccd86 ko 2021-09-01 ma karobar rakam (turnover) kati thiyo?	nccd86 ko 2021-09-01 ma karobar rakam 25,177.50 thiyo.
Trans.	nil ko 2021-09-01 ma karobar sankhya (transactions) kati thiyo?	nil ko 2021-09-01 ma karobar sankhya 395 thiyo.
VWAP	igi ko 2021-09-01 ma vwap mulya kati thiyo?	igi ko 2021-09-01 ma vwap mulya 1,192.39 thiyo.
Diff	pcbl ko 2021-09-01 ma mulya pharak (diff) kati thiyo?	pcbl ko 2021-09-01 ma mulya pharak 3.50 thiyo.
Range	rhpc ko 2021-09-01 ma mulya dayara (range) kati thiyo?	rhpc ko 2021-09-01 ma mulya dayara 62 thiyo.
Diff %	sbi ko 2021-09-01 ma mulya pariwartana pratishat (diff %) kati thiyo?	sbi ko 2021-09-01 ma mulya pariwartana pratishat -0.78% thiyo.
Range %	shpc ko 2021-09-01 ma mulya dayara pratishat (range %) kati thiyo?	shpc ko 2021-09-01 ma mulya dayara pratishat 4.68% thiyo.
VWAP %	smfdb ko 2021-09-01 ma vwap pratishat kati thiyo?	smfdb ko 2021-09-01 ma vwap pratishat 0.71% thiyo.
Close - LTP	lsl ko 2025-03-27 ma close ra ltp bichko pharak kati thiyo?	lsl ko 2025-03-27 ma close ra ltp bichko pharak 0.50 thiyo.
Close - LTP %	mhcl ko 2025-03-27 ma close-ltp pratishat pharak kati thiyo?	mhcl ko 2025-03-27 ma close-ltp pratishat pharak -0.02% thiyo.
120 Days	bpcl ko 2021-11-03 ma 120 dinko ausat mulya kati thiyo?	bpcl ko 2021-11-03 ma 120 dinko ausat mulya 527.27 thiyo.
180 Days	ddbl ko 2021-11-03 ma 180 dinko ausat mulya kati thiyo?	ddbl ko 2021-11-03 ma 180 dinko ausat mulya 1,728.31 thiyo.
52 Weeks High	gimes1 ko 2021-11-03 ma 52 haptako uchcha mulya kati thiyo?	gimes1 ko 2021-11-03 ma 52 haptako uchcha mulya 25.09 thiyo.
52 Weeks Low	icfcd83 ko 2021-11-03 ma 52 haptako nyuna mulya kati thiyo?	icfcd83 ko 2021-11-03 ma 52 haptako nyuna mulya 1,080 thiyo.
Conf.	uslb ko 2021-09-01 ma wishwas suchak (conf.) kati thiyo?	uslb ko 2021-09-01 ma wishwas suchak 45.63 thiyo.

---

11. Intended Uses

Use	Fit	Why
Fine-tuning a model on Romanized Nepali finance vocabulary	Good	22 distinct finance terms in consistent Romanized spelling
Teaching a model to copy entities, dates and numbers exactly	Good	The answer repeats ticker, date and metric, then gives one number
Numeric-format learning (commas, decimals, negatives, percent)	Good	Five value kinds with several formats
Evaluating exact-match accuracy on short factual answers	Good	One correct number per question, easy to parse with a regular expression
Building a retrieval or tool-calling benchmark for NEPSE data	Good	Questions are fully determined by (ticker, date, metric)
Open-domain Nepali chat	Poor	Only 21 rigid templates, no free-form language
Financial advice, forecasting or reasoning	Not suitable	No analysis, advice or multi-step reasoning is present
Learning true Nepali (Devanagari) text generation	Not suitable	All text is Latin script (see Section 12)
Model memorisation of prices	Be careful	A model cannot know a price without the source table; use retrieval or tools

---

12. Known Limitations and Metadata Notes

#	Observation	Impact	Suggestion
1	The metadata declares script: Deva, content_script: देवनागरी, content_language: नेपाली and english_content_allowed: false, but every question and answer is Romanized Latin text with zero Devanagari characters	Anything that filters or routes by these tags will mislabel the data	Treat the real script as Latin; override the tags if you publish this dataset
2	The behavior_definition is written in Devanagari but describes producing "pure Nepali" output, while the answers are Romanized	Mismatch between the label text and the content	Same as above
3	Only one behaviour and one question type exist	No instruction diversity	Mix with other datasets for general assistants
4	Each target has exactly one question wording (21 templates in total)	Models can overfit to the templates and fail on paraphrased questions	Add paraphrases or augment the questions before training
5	The dataset has no train / validation / test split (source_split is train for all rows)	Random splits can leak (same ticker, same date, different target)	Use the date-based split in Section 13
6	The file contains dates on all seven weekdays (see Section 7.3), including Fridays and Saturdays	The dates may be the source's record dates rather than guaranteed trading days; this was not verified against an exchange calendar	Check against the source before using the dates for market analysis
7	Extreme values exist, for example VWAP % down to -400.00 and Prev. Close equal to 0 in 2 records	Possible source-data anomalies	Inspect or clip outliers if you train a numeric model
8	The currency and units are never stated in the text	Ambiguity for readers outside the Nepali market	Document the units in your own downstream card
9	Values come from the scrape as-is and were not independently re-verified against the exchange	Source errors would be inherited	Do not use for trading decisions
10	Conf. is translated here as "confidence indicator" from the Romanized phrase wishwas suchak; its exact definition is not given in the data	Interpretation risk	Check the source repository for the column definition
11	Targets LTP, Close - LTP and Close - LTP % exist only from 2025-03-26 onwards	Uneven target availability over time	Account for it in time-based analysis
12	metadata_json is a JSON-encoded string rather than a nested object	Extra parsing step, and some loaders will not expose its keys as columns	Use the code in Section 9

---

13. Suggested Train / Validation / Test Splits

The file ships as one block. The recommended way to split it is by date, so that the same ticker-date pair never appears in two splits.

Split	Date range	Records	Share
train	before 2026-01-01	26,282	86.17%
validation	2026-01-01 to 2026-04-30	2,496	8.18%
test	2026-05-01 onwards	1,722	5.65%

(Counts are for the cut points 2026-01-01 and 2026-05-01, matching the code in Section 9.6.)

---

14. License and Citation

Item	Value
License tag in the records	Apache-2.0 (license_tier: permissive)
Where the license comes from	The upstream repository sbmagar13/sharesansar_datascrape
Recommendation	Confirm the upstream license and the terms of use of the original website data before redistributing

Suggested citation stub (edit it to your own details):

@dataset{sharesansar_romanized_qa_30k,
  title        = {ShareSansar Romanized Nepali Stock-Market QA (30.5k)},
  year         = {2026},
  note         = {Template-generated factual QA built from sbmagar13/sharesansar_datascrape},
  howpublished = {JSONL, ShareGPT-style conversations}
}