sidddd625/adaption-banking-and-digital-payments

This dataset is a remastered version prepared using Adaption's Adaptive Data platform.

adaption-Banking and Digital Payments

This dataset contains question-answer pairs focused on Indian personal finance, digital payment security, and banking regulations. The content covers UPI usage, fraud prevention, RBI guidelines, and alternative payment methods for diverse user groups including seniors and feature phone users. Each entry provides actionable advice grounded in official mandates from entities like NPCI and UIDAI. Made by team 10_minutes as a part of our solution focusing on financial inclusion in India for HackIndia hackathon with the help of adaptive dataset creation using Adaption.

Dataset size

There are 2,631 data points in this dataset. This is an instruction tuning dataset.

Quality of Remastered Dataset

The final quality is A, with a relative quality improvement of 30.0%.

Domain

• Personal-finance (70%)
• Legal (24%)
• Technology (4%)

Language

• English (98%)
• German (2%)

Tone

• Helpful (48%)
• Cautious (16%)
• Practical (14%)

Evaluation Results

• Quality Gains:

  

• Grade Improvement:

  

• Percentile Chart:

  

🏦 Banking & Digital Payments — Fine-Tuning Dataset

Domain: Banking & Digital Payments
Format: JSONL (JSON Lines)
Size: ~54.6 MB
Languages: 14 Indian languages
Processed by: Adaption

---

Overview

India's digital payments revolution — driven by UPI, BHIM, mobile banking, and net banking — has brought hundreds of millions of first-time banking users online. But with access comes confusion: fake OTP requests, failed transactions, blocked accounts, and misunderstood limits leave users stranded. This dataset captures the full spectrum of real-world struggles and questions that Indian banking customers face every day.

Each record in this dataset is a structured financial Q&A grounded in an authentic user scenario — from a homemaker who clicked a fake SBI KYC alert, to a vendor trying to reconcile a UPI payment that settled but didn't reflect. The prompts are real queries; the completions are expert-grade, step-by-step guidance aligned with RBI and NPCI guidelines.

---

Dataset Schema

Each record is a JSON object with the following 12 fields, confirmed from the actual data:

Field	Type	Description
user_query	string	The raw, original user question — exactly as sourced from forums, community Q&As, or curated inputs
enhanced_prompt	string	The Adaption-rephrased version of the query: restructured for clarity, enriched with context, and optimized for model instruction-following
combined_completion	string	A baseline merged completion combining the original answer guidance, actions, and learning outcome — used as the pre-enhancement reference
enhanced_completion	string	The final Adaption-processed expert answer: detailed, step-by-step, and fully localized in the user's language — this is the primary training target
answer_guidance	string	Concise English factual summary grounding the answer in official policy — used for training alignment and evaluation
actions_suggestions_next_step	string	Numbered, practical action plan the user should follow after reading the answer
domain_category	string	Top-level domain — always "Banking & Digital Payments" in this file (1 unique value)
subdomain	string	Granular topic within the domain — 28 distinct subdomains in this file (e.g., "UPI Fraud Prevention", "Net Banking")
userprofile	string	Demographics, occupation, digital literacy level, and situational context of the user
learning_outcome	string	The financial capability or awareness the user is meant to gain (188 distinct values across records)
source	string	Authoritative references — RBI circulars, CERT-In bulletins, TRAI regulations, NPCI guidelines, bank security pages
feedback	string	Post-interaction outcome — what the user actually did and whether the issue was resolved (188 distinct values)

The Training Pipeline: How Fields Relate

 [user_query]  →  Adaption Prompt Rephrase  →  [enhanced_prompt]
      ↓                                               ↓
 [raw sources]  →  Adaption Data Processing  →  [enhanced_completion]
                                                        ↓
                                          [combined_completition]  ← merged baseline reference

For fine-tuning, the primary training pair is enhanced_prompt → enhanced_completion. The user_query and combined_completition fields preserve the pre-Adaption originals for auditability, ablation studies, and quality comparison.

Example Record (Actual Fields)

{
  "user_query": "I got an SMS from SBI-ALERTT asking me to click a link to update KYC...",
  "enhanced_prompt": "[Metadata-enriched, rephrased version for model training]",
  "combined_completition": "[Baseline merged answer in the user's language]",
  "enhanced_completion": "[Full step-by-step localized expert response — the training target]",
  "userprofile": "Homemaker who received a fake SBI KYC alert SMS and clicked the link",
  "subdomain": "UPI Fraud Prevention",
  "domain_category": "Banking & Digital Payments",
  "answer_guidance": "Yes, this is a phishing SMS (smishing). Fraudsters spoof SMS sender IDs...",
  "actions_suggestions_next_step": "1. Change internet banking password immediately via the official app...",
  "source": "CERT-In Phishing Alert Advisories, RBI Anti-Phishing Guidelines, TRAI Regulations",
  "learning_outcome": "User identifies smishing attacks through suspicious sender IDs and urgency language",
  "feedback": "Credentials changed in time. No financial loss reported. Complaint filed with cyber cell."
}

---

Subdomains Covered

This dataset spans the full lifecycle of a banking customer's digital journey:

• UPI & BHIM — payment failures, wrong credits, merchant disputes, daily limits
• Net Banking & Mobile Banking — login issues, blocked accounts, 2FA, MPIN resets
• NEFT / RTGS / IMPS — transaction timelines, failed transfers, beneficiary management
• Debit & Credit Cards — international usage, chargeback, EMI conversion
• Account Operations — KYC updates, nomination, joint accounts, account closure
• UPI Fraud Prevention — smishing, vishing, fake payment links, QR scams
• Digital Wallet Regulations — Paytm, PhonePe, Google Pay — limits and compliance

---

Sources & Grounding

Every completion and answer_guidance is backed by authoritative sources embedded in the source field. These allow fine-tuned agents to trace answers back to primary documents and — when deployed with web search or RAG — fetch updated policy text directly. Sources in this dataset include:

• RBI Circulars — fraud alerts, KYC guidelines, transaction dispute timelines
• NPCI Documentation — UPI operational guidelines, dispute resolution procedures
• CERT-In Advisories — phishing attack bulletins specific to Indian banking
• TRAI Regulations — unsolicited commercial communication rules
• Bank Official Policies — SBI, HDFC, ICICI, Axis security and operations pages

---

User Personas

The dataset was built to reflect India's diverse banking population:

• 🏡 Rural women — first-time Jan-Dhan account holders navigating mobile banking
• 🛒 Market vendors — using UPI QR codes daily, encountering settlement delays
• 👩‍💼 Urban professionals — managing salary accounts, international cards, and SIPs
• 🧓 Senior citizens — confused by 2FA and OTP-based banking transitions
• 🏗️ Migrant workers — sending remittances home via IMPS, dealing with KYC gaps

---

Localization — Why, Not Just Translation

This dataset was localized into 14 Indian languages — not mechanically translated.

The 14 languages: English · Hindi · Marathi · Gujarati · Bengali · Tamil · Telugu · Kannada · Konkani · Bhojpuri · Maithili · Odia · Malayalam · Punjabi

Why localization matters here:
Banking in India is deeply tied to institutional names — "SBI YONO", "BHIM UPI", "Jan-Dhan", "PMJDY" — and regulatory bodies like "RBI", "CERT-In", and "TRAI". These cannot be translated; they are proper nouns that carry meaning only as-is. A translated answer that renders "BHIM" or "KYC" into a regional script equivalent would confuse the very user it tries to help.

Localization preserves these anchors exactly as they appear in official communications while adapting the explanation, tone, and phrasing around them to be natural and culturally resonant in each language. A farming household in rural Odisha gets the same factual accuracy as an IT professional in Bengaluru — just in a language and register they actually speak at home.

This approach is central to the project's mission: financial inclusion through language, not just language access.

---

How Adaption Processed This Data

Raw scraped data, community Q&As, and hand-curated queries went through Adaption's full data pipeline:

1. Prompt Rephrase — each user query was rewritten for clarity and model instruction-following quality, while preserving the original intent and colloquial tone
2. Prompt Deduplication — near-duplicate questions (same issue, slightly different phrasing) were detected and consolidated, keeping the highest-quality variant
3. Prompt Metadata Injection — structured metadata (userprofile, subdomain, learning_outcome) was injected into each prompt to give the model richer context about who is asking and why

Adaption's internal benchmarks show an average 82% improvement in data quality after this pipeline — measurable in answer accuracy, relevance, and instruction-following.

---

Training Specialized Agents

This dataset is designed for fine-tuning LLMs into domain-expert agents for the Indian banking context. A model trained on this data can:

• Understand the nuance between a UPI dispute and a bank fraud complaint
• Guide users through step-by-step remediation in their native language
• Cite the right regulatory body (RBI vs. NPCI vs. CERT-In) for each issue type
• Use the source field to perform RAG lookups for real-time policy verification
• Serve users across literacy levels, from urban professionals to first-time smartphone users

---

Part of a Larger Collection

This dataset is one of five domain-specific files. The combined master dataset is available as india_finance_qa.jsonl.

Dataset	Domain
Banking & Digital Payments ← you are here	UPI, net banking, cards, accounts
Credit & Borrowing	Loans, CIBIL, EMI, credit cards
Fraud & Cyber Safety	Scams, phishing, account takeovers
Savings & Insurance	FD, mutual funds, PMSBY, life insurance
Government Schemes	PMJDY, PM Kisan, subsidies, eligibility