sabin1234/Nepali_Commercial_Bank_Financial_Indicators_QA_Dataset

language:

• —ne license: cc-by-4.0 task_categories:
• —question-answering
• —text-generation tags:
• —nepali
• —nepal
• —banking
• —finance
• —commercial-banks
• —grounded-qa
• —sharegpt
• —instruction-tuning size_categories:
• —1K<n<10K ---

Nepali Commercial Bank Financial Indicators — Grounded QA Dataset

A Nepali-language, grounded single-metric question–answering dataset built from the quarterly Key Financial Indicators of Commercial Banks published for Nepal's commercial banking sector. Each example is a single-turn human↔assistant conversation (ShareGPT / Hermes style) in which a question about one specific bank, one specific quarter, and one specific financial metric is answered with the exact value drawn from the source data.

This dataset is designed for instruction fine-tuning and retrieval/grounding evaluation of LLMs on Nepali financial-domain factual recall.

1. Dataset Summary

2. License

• —License: CC BY 4.0 (Creative Commons Attribution 4.0 International)
• —Declared per-record in the source_provenance.metadata.license_id field — 100% of the 7,421 records carry "CC BY 4.0".
• —You are free to share (copy and redistribute) and adapt (remix, transform, build upon) the dataset for any purpose, including commercially, provided you give appropriate credit to the original data source (Commercial Bank Financial Indicators / Nepal's commercial bank quarterly disclosures), link to the license, and indicate if changes were made.
• —No warranty is provided; the data is distributed "as-is." Verify current published figures against the primary regulatory/bank source before using this data for real financial decision-making.
• —Full license text: https://creativecommons.org/licenses/by/4.0/

3. Dataset Structure

Each line in the .jsonl file is a single JSON object shaped like this:

{
  "id": "grounded-qa-000001",
  "conversations": [
    {"from": "human", "value": "२०७७(१२) मा नेपाल बैंक लिमिटेडका लागि रिपोर्ट गरिएको सोल्भेन्सी कोर क्यापिटल कति थियो?"},
    {"from": "gpt",   "value": "२०७७(१२) मा नेपाल बैंक लिमिटेडका लागि रिपोर्ट गरिएको सोल्भेन्सी कोर क्यापिटल २३२२० थियो।"}
  ],
  "source_provenance": {
    "id": "provenance-000001",
    "grounding": {
      "bank": "Nepal Bank Ltd.",
      "period": "2077(12)",
      "metric": "Solvency Core Capital",
      "value": "23220"
    },
    "metadata": {
      "dataset_name": "Nepali Commercial Bank Financial Indicators ShareGPT Dataset",
      "dataset_domain": "Nepalese Commercial Banking and Financial Indicators",
      "resource_name": "Key Financial Indicators of Commercial Banks (Quarterly)",
      "source_name": "Commercial Bank Financial Indicators",
      "license_id": "CC BY 4.0",
      "datatype": "CSV",
      "source_column": "Solvency Core Capital",
      "model_name": "gemini-3.5-flash-lite, gemini-3.6-flash"
    },
    "behavior": "grounded_single_metric_qa",
    "source_instruction": "What was the solvency core capital reported for Nepal Bank Ltd. in 2077(12)?",
    "dataset_name": "Nepali Commercial Bank Financial Indicators ShareGPT Dataset",
    "dataset_split": "train",
    "row_index": 0,
    "resource_id": "resource-000001",
    "source_task_id": "grounded-qa-000001"
  },
  "input_format": "hermes-instruction-response",
  "source_task_id": "grounded-qa-000001",
  "model_name": "gemini-3.5-flash-lite, gemini-3.6-flash"
}

Field reference

4. Coverage: Banks & Periods

58 unique bank-name strings appear in the grounding.bank field. Note this count includes:

• —Aggregate rows: Sub Total (560 rows, 7.55%) and Grand Total (300 rows, 4.04%), which are not individual banks but sector-level rollups present in the source CSV.
• —Multiple naming variants for the same institution across different reporting periods (e.g., Nepal Bank Ltd. / Nepal Bank Limited, NIC Asia Bank Ltd. / NIC ASIA Bank Limited, Global IME Bank Ltd. / GlobalIME Bank Ltd. / Global IME Bank Limited), reflecting inconsistent naming in the original quarterly disclosures rather than distinct banks. If you need a clean bank list, you should normalize these strings first (see §7).

Individual (non-total) banks each account for roughly 0.2%–2.3% of the dataset, e.g. Nepal Bank Ltd. (~2.14%), Nabil Bank Ltd. (~2.25%), Rastriya Banijya Bank Ltd. (~2.25%), down to smaller counts for banks that merged, changed names, or exited the market partway through the period.

18 reporting quarters (B.S.) are covered, from 2077(12) through 2082(03) — i.e. roughly mid-2020 to mid-2025 in the Gregorian calendar — with between 362 and 485 examples per quarter.

5. Behavior / Task Diversity

The dataset's source_provenance.behavior field records the QA-generation template used. Distribution across the 7,421 total examples:

➡️ This dataset is single-behavior: every example follows the same template — a factual, single-metric question grounded in one (bank, period, metric) triple, answered with the exact source value. There is no multi-turn reasoning, comparison-across-banks, comparison-across-periods, calculation, or summarization behavior present in this file. If diverse task types are important for your use case, this dataset should be combined with other behavior templates (e.g., trend/comparison QA) generated from the same source CSV.

Diversity that does exist (within the single behavior)

Metric distribution (17 metrics, all ~5.5%–6.1% of the data — well balanced):

6. How to Load This Dataset

Option A — 🤗 datasets library (recommended for HF Hub)

from datasets import load_dataset

# If uploaded to the Hugging Face Hub under your namespace:
ds = load_dataset("your-username/nepali-commercial-bank-financial-indicators-qa", split="train")

# If loading the raw local/uploaded JSONL file directly:
ds = load_dataset(
    "json",
    data_files="nepali_commercial_bank_financial_indicators.jsonl",
    split="train"
)

print(ds)
print(ds[0])

Expected output shape:

Dataset({
    features: ['id', 'conversations', 'source_provenance', 'input_format', 'source_task_id', 'model_name'],
    num_rows: 7421
})

Option B — Plain Python (jsonlines / manual read)

import json

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

print(f"Loaded {len(records)} examples")
print(records[0]["conversations"][0]["value"])  # Nepali question
print(records[0]["conversations"][1]["value"])  # Nepali answer

Option C — pandas (for quick tabular analysis)

import json
import pandas as pd

rows = []
with open("nepali_commercial_bank_financial_indicators.jsonl", encoding="utf-8") as f:
    for line in f:
        obj = json.loads(line)
        g = obj["source_provenance"]["grounding"]
        rows.append({
            "id": obj["id"],
            "question": obj["conversations"][0]["value"],
            "answer": obj["conversations"][1]["value"],
            "bank": g["bank"],
            "period": g["period"],
            "metric": g["metric"],
            "value": g["value"],
        })

df = pd.DataFrame(rows)
print(df.head())
print(df["metric"].value_counts(normalize=True) * 100)

Option D — Fine-tuning frameworks (Axolotl / FastChat ShareGPT format)

Because conversations already uses the from/value ShareGPT schema, this file can be pointed to directly as a sharegpt-type dataset in most fine-tuning configs (e.g. Axolotl's type: sharegpt), with no reformatting needed beyond optionally stripping the source_provenance metadata for training.

7. Suggested Preprocessing

If you plan to use this for analysis or training where clean entity labels matter, consider:

1. 1.Normalizing bank names — merge variants like Nepal Bank Ltd. / Nepal Bank Limited / Nepal Bank Limite into one canonical name.
2. 2.Deciding whether to keep `Sub Total` / `Grand Total` rows — these are sector aggregates, not individual banks, and may need to be excluded or handled separately depending on your downstream task.
3. 3.Converting B.S. periods to A.D. dates if you need Gregorian-calendar alignment for time-series work.

8. Source & Provenance

• —Source data: "Key Financial Indicators of Commercial Banks (Quarterly)" — a CSV resource under the dataset "Commercial Bank Financial Indicators," within the broader domain of Nepalese Commercial Banking and Financial Indicators.
• —Generation pipeline: English seed instructions (source_instruction) were used to generate grounded Nepali question–answer pairs via `gemini-3.5-flash-lite` and `gemini-3.6-flash`, with each answer traceable back to a specific (bank, period, metric, value) cell in the source CSV via the source_provenance field — enabling full auditability of every generated answer against its ground-truth source value.
• —File name (as provided): _openhermes_nepali_cleaned_v3__Key_Financial_indicators_of_Commerical_banks__quaterly_.jsonl
• —Version: v3 (per file naming convention — "cleaned_v3")

9. Intended Use

• —Fine-tuning Nepali-language LLMs / chat assistants for grounded, factual financial Q&A.
• —Benchmarking hallucination/grounding accuracy of LLMs on structured domain data.
• —Evaluation of Nepali numeral, Devanagari script, and B.S.-calendar handling in generative models.

10. Limitations

• —Single behavior type only (see §5) — not suitable alone for teaching comparison, aggregation, or multi-metric reasoning.
• —Bank-name inconsistency across periods requires normalization for clean entity-level analysis (§4, §7).
• —Figures reflect a point-in-time snapshot of quarterly disclosures generated via an LLM pipeline; always cross-check against the original regulator/bank-published source before relying on the values for real decisions.
• —Not reviewed for translation quality by a native financial-domain human annotator — treat generated Nepali phrasing as unvalidated at the individual-example level.

Citation

If you use this dataset, please attribute per its CC BY 4.0 license:

Nepali Commercial Bank Financial Indicators ShareGPT Dataset,
derived from "Key Financial Indicators of Commercial Banks (Quarterly)",
licensed under CC BY 4.0.