sabin1234/Failed_Bank_Dataset

🏦 Failed Bank Dataset (Nepali) — ShareGPT Format

4,528 Nepali (नेपाली) instruction-following conversations built from real U.S. failed-bank records — CC0 Public Domain.

If America's bank failures had to explain themselves... in Nepali... this is that dataset. Yes, really. 😄

---

📌 TL;DR

File	Failed_Bank_Dataset_2_Nepali_ShareGPT_4528_CC0_FINAL_SUDDHA_FINAL.jsonl
Format	JSONL, ShareGPT-style (conversations: [human, gpt])
Total examples	4,528
Unique failed banks covered	566
Language	Nepali (ne / npi), Devanagari script
License	CC0: Public Domain (do whatever you want, no attribution needed)
Turns per conversation	Always 2 (1 human + 1 gpt) — clean single-turn QA
File size	~7.8 MB

This is a fully synthetic-templated, fact-grounded QA dataset: real closed-bank data (source city, state, acquiring institution, closing date, estimated loss, total assets, total deposits) turned into natural-sounding Nepali question–answer pairs. Think of it as "FDIC failed bank list, but it speaks Nepali and answers politely."

---

🎨 Design — How the Dataset is Built

Every one of the 566 failed banks gets asked about 8 different angles of its record. 566 × 8 = 4,528 — the math checks out perfectly, so no bank was left un-interrogated. 🕵️

1 failed bank record  ×  8 question angles  =  8 conversations per bank
566 banks             ×  8 angles           =  4,528 total conversations

The 8 "angles" (called generation_sub_domain in the metadata) are perfectly evenly distributed — 566 examples each, no exceptions:

#	Sub-domain (Nepali)	What it asks about	Count	Share
1	स्थान (Location)	Which city the bank was listed in	566	12.5%
2	राज्य (State)	Which U.S. state the city belongs to	566	12.5%
3	अधिग्रहण संस्था (Acquirer)	Which institution acquired the failed bank	566	12.5%
4	बन्द मिति (Closing Date)	When the bank officially closed	566	12.5%
5	अनुमानित नोक्सानी (Estimated Loss)	Estimated financial loss from the failure	566	12.5%
6	कुल सम्पत्ति (Total Assets)	Total assets on record	566	12.5%
7	कुल निक्षेप (Total Deposits)	Total deposits on record	566	12.5%
8	समग्र विवरण (Full Summary)	Everything above, combined into one answer	566	12.5%

Basically: 7 fact-lookup questions + 1 "give me the whole story" summary question, per bank. Every distribution below is this same clean 8-way split underneath — behavior, question type, and domain are all single-mode by design (see below), so the real variety in this dataset comes from how each of these 8 angles is phrased, not from category diversity.

---

📊 Complete Distribution Breakdown

1. Behavior distribution (behavior)

There is exactly one behavior type in this dataset — every single response follows it:

Behavior	Definition	Count	Share
प्रत्यक्ष तथ्यात्मक उत्तर दिने (Give a direct factual answer)	Present a clear answer in pure Nepali while preserving the original fact	4,528	100%

No hedging, no "I'm not sure," no roleplay, no refusals. It's a straight-shooting factual-QA dataset — the model equivalent of a bank teller who just wants to give you your balance and move on to the next customer.

2. Question type distribution (question_type)

Question Type	Count	Share
तथ्यमा आधारित प्रश्न (Fact-based question)	4,528	100%

Again — single mode, 100%. This is intentional: it's a grounded-fact QA corpus, not an opinion/creative/reasoning dataset. Great for factual-recall fine-tuning, weak for teaching a model to argue or brainstorm.

3. Domain / task-type distribution

Field	Value	Count	Share
generation_domain	बैंकिङ (Banking)	4,528	100%
generation_category	असफल बैंक अभिलेख (Failed Bank Records)	4,528	100%
task_type	निर्देशन पालन (Instruction Following)	4,528	100%
generation_type	वास्तविक (Real/Factual)	4,528	100%
condition	वास्तविक (Real)	4,528	100%
language	ne (Nepali)	4,528	100%
script	Deva (Devanagari)	4,528	100%
license	CC0: Public Domain	4,528	100%
english_content_allowed	false	4,528	100%

In short: this dataset is 100% narrow and 100% deep. One domain (banking), one category (bank failures), one language, one script, one license — but 566 real-world entities examined from 8 angles each. It's less a "buffet" dataset and more a "single perfectly-cooked dish, 566 times." 🍛

4. Conversation structure distribution

Turns	Count	Share
2 (human → gpt)	4,528	100%

No multi-turn chains, no follow-up corrections — every example is a clean single question, single answer.

5. Length distribution (rough sense of verbosity)

	Human (question)	GPT (answer)
Avg. words	~11.1	~13.7
Avg. characters	~67	~86
Shortest	6 words	6 words
Longest	34 words	47 words (usually the "full summary" answers)

Questions are short and punchy; answers ("समग्र विवरण" ones especially) can run a bit longer since they stitch all 6 facts into one paragraph.

6. Unique entity coverage

• 566 unique failed banks/institutions, each with a unique source_row_id.
• Every single bank has exactly 8 conversations tied to it — no bank is over- or under-represented. Very tidy bookkeeping, appropriately, for a dataset about banks.

---

🧾 Field Reference

Each JSONL line is one conversation object:

{
  "id": "failed_bank_faq_000001",
  "conversations": [
    {"from": "human", "value": "कुन सहरमा फर्स्ट रिपब्लिक बैंक को शाखा/संस्था सूचीकृत छ, र त्यस सहरसँग कुन राज्य जोडिएको छ?"},
    {"from": "gpt",   "value": "फर्स्ट रिपब्लिक बैंक को सूचीबद्ध सहर सान फ्रान्सिस्को हो। यो स्थान क्यालिफोर्निया राज्यसँग सम्बन्धित छ।"}
  ],
  "source": "Failed Bank Dataset",
  "source_name": "failed_bank_dataset_2",
  "source_repo": "Failed Bank Dataset",
  "source_config": "default",
  "source_split": "train",
  "source_revision": "v1",
  "source_row_id": "failed_bank_row_0001",
  "language": "ne",
  "language_code": "npi",
  "script": "Deva",
  "license": "CC0: Public Domain",
  "license_tier": "permissive",
  "task_type": "निर्देशन पालन",
  "generation_type": "वास्तविक",
  "condition": "वास्तविक",
  "url": "",
  "metadata_json": "{...stringified JSON, see below...}"
}

metadata_json is a stringified JSON blob (you need to json.loads() it separately) containing:

Key	Meaning	Example value
generation_domain	Broad domain	बैंकिङ (Banking)
generation_category	Category within domain	असफल बैंक अभिलेख (Failed Bank Records)
generation_sub_domain	The specific fact-angle asked	स्थान / राज्य / बन्द मिति / ...
behavior	Expected model behavior	प्रत्यक्ष तथ्यात्मक उत्तर दिने
behavior_definition	What that behavior means	(see Section 1 above)
question_type	Type of question	तथ्यमा आधारित प्रश्न
content_language	Content language (human-readable)	नेपाली
content_script	Script used	देवनागरी
english_content_allowed	Whether English is allowed in content	false

---

💻 How to Load This Dataset (Demo Code)

Option 1 — Plain Python (no dependencies)

import json

path = "Failed_Bank_Dataset_2_Nepali_ShareGPT_4528_CC0_FINAL_SUDDHA_FINAL.jsonl"

records = []
with open(path, "r", encoding="utf-8") as f:
    for line in f:
        line = line.strip()
        if not line:
            continue
        records.append(json.loads(line))

print(f"Loaded {len(records)} conversations")

# Peek at the first example
example = records[0]
human_msg = example["conversations"][0]["value"]
gpt_msg = example["conversations"][1]["value"]
print("Q:", human_msg)
print("A:", gpt_msg)

# metadata_json is a *string* — parse it separately
meta = json.loads(example["metadata_json"])
print("Sub-domain:", meta["generation_sub_domain"])

Option 2 — Pandas (for filtering / analysis)

import json
import pandas as pd

path = "Failed_Bank_Dataset_2_Nepali_ShareGPT_4528_CC0_FINAL_SUDDHA_FINAL.jsonl"

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

# Expand the metadata_json string column into real columns
meta_df = pd.json_normalize(df["metadata_json"].apply(json.loads))
df = pd.concat([df.drop(columns=["metadata_json"]), meta_df], axis=1)

# Extract question/answer text into their own columns
df["question"] = df["conversations"].apply(lambda c: c[0]["value"])
df["answer"]   = df["conversations"].apply(lambda c: c[1]["value"])

print(df.shape)
print(df["generation_sub_domain"].value_counts())

# Example: get only the "full summary" QA pairs
summaries = df[df["generation_sub_domain"] == "समग्र विवरण"]
print(summaries[["question", "answer"]].head())

Option 3 — Hugging Face datasets library

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files="Failed_Bank_Dataset_2_Nepali_ShareGPT_4528_CC0_FINAL_SUDDHA_FINAL.jsonl",
    split="train",
)

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

Option 4 — Format for chat fine-tuning (e.g. OpenAI / Llama-style messages)

import json

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

path = "Failed_Bank_Dataset_2_Nepali_ShareGPT_4528_CC0_FINAL_SUDDHA_FINAL.jsonl"
with open(path, encoding="utf-8") as f, open("chat_format.jsonl", "w", encoding="utf-8") as out:
    for line in f:
        record = json.loads(line)
        out.write(json.dumps(to_chat_format(record), ensure_ascii=False) + "\n")

print("Converted to standard chat/messages format ✅")

---

✅ Good Use Cases

• Fine-tuning / instruction-tuning small Nepali LLMs on factual QA behavior
• Nepali language modeling with real-world numeric & named-entity data (dates, dollar amounts, institution names)
• Benchmarking a model's ability to answer grounded factual questions in Devanagari script
• Testing/eval sets for "does the model hallucinate numbers" (since every fact is traceable to a real closed-bank record)

⚠️ Limitations / Things to Know Before You Use It

• Not diverse in behavior or question type — 100% factual QA, single-turn, no dialogue/chit-chat/reasoning/refusal examples. Don't expect this alone to teach a model personality or multi-turn skills.
• Narrow domain — banking/failed-institutions only. Great for domain fine-tuning, not a general-purpose Nepali dataset on its own.
• Templated phrasing — questions are naturally varied per sub-domain, but the underlying pattern (566 banks × 8 fixed angles) repeats structurally across the whole file.
• Some fields (e.g., "estimated loss") are legitimately marked as "not specified" (उल्लेख गरिएको छैन) in the original source records — that's expected, not missing data.

---

📜 License

CC0: Public Domain. No rights reserved. Use it, remix it, sell models trained on it, print it out and use it as wallpaper — nobody's stopping you. 🎉

---

README generated for the Failed Bank Dataset (Nepali, ShareGPT format, 4,528 examples, CC0).