RakeshMadasani/banking-finance-mistral-qlora

This model is a QLoRA fine-tuned version of mistralai/Mistral-7B-Instruct-v0.3 adapted for banking, finance, AML/KYC, compliance, and regulatory question answering.

It was trained on a custom banking and finance instruction dataset built from official regulatory and domain-specific material, then published as part of a 3-project GenAI portfolio:

• Banking RAG Assistant
• Banking & Finance QA Dataset
• Banking Finance QLoRA Fine-Tuned Model

This model is intended for domain-style question answering across topics such as:

• AML and KYC
• FDIC and deposit insurance
• Basel III and capital adequacy
• RBI and Indian banking regulations
• SAR / CTR / compliance reporting
• credit risk and financial concepts

The goal of the project was to improve domain-specific response quality compared with the base model while keeping training efficient through PEFT and QLoRA.

The model was fine-tuned on:

• RakeshMadasani/banking-finance-qa-dataset

Dataset summary:

• 3,002 instruction-response pairs
• Alpaca-style format
• English language
• Banking and finance domain
• Topics include AML, KYC, FDIC, Basel, RBI, SAR, CTR, and compliance

Fine-tuning approach:

• QLoRA
• PEFT / LoRA adapters
• 4-bit quantization
• Mistral-7B-Instruct base model

Training setup:

• Base model: mistralai/Mistral-7B-Instruct-v0.3
• Method: QLoRA
• LoRA rank: 16
• LoRA alpha: 32
• Dropout: 0.05
• Max sequence length: 384-512 during experimentation
• Platform: GPU notebook environment
• Objective: domain adaptation for banking and compliance question answering

Example questions:

• What is the FDIC deposit insurance limit in the United States?
• What are the three stages of money laundering?
• What documents are required for KYC of an individual in India?
• What is a Suspicious Activity Report and when must it be filed?
• What is the minimum capital adequacy ratio under Basel III in India?

This model is intended for:

• educational demonstrations
• portfolio projects
• domain adaptation experiments
• banking/compliance QA prototyping

This model:

• is not a substitute for legal, compliance, or financial advice
• may still produce incomplete or incorrect answers
• should not be used for real regulatory decision-making without verification
• performs best when prompts are clearly phrased and domain-specific

The fine-tuning workflow was validated through: - successful end-to-end training completion - sample-based qualitative evaluation on banking and compliance questions - comparison against expected domain answers on topics such as FDIC insurance, AML stages, and Basel / compliance concepts

license: apache-2.0 task_categories: - text-generation - question-answering language: - en tags: - banking - finance - aml - kyc - compliance - instruction-tuning - alpaca - qlora pretty_name: Banking & Finance QA Dataset size_categories: - 1K<n<10K

Banking & Finance QA Dataset

This dataset is a custom instruction-response dataset created for banking, finance, AML/KYC, compliance, and regulatory question answering.

It was built as part of a 3-project GenAI portfolio:

• Banking RAG Assistant

• Banking & Finance QA Dataset

• Banking Finance QLoRA Fine-Tuned Model

• Total samples: 3,002

• Format: Alpaca-style instruction / input / output

• Language: English

• Domain: Banking and Finance

• Primary use case: Instruction tuning / QLoRA fine-tuning

Each record follows this structure:

{
  "instruction": "What is the FDIC deposit insurance limit in the United States?",
  "input": "",
  "output": "The FDIC deposit insurance limit is $250,000 per depositor, per insured bank, per account category."
}