🏦 Fine-tuning LLMs for RBI Regulatory Q&A

Fine-tuning Qwen 2.5 3B on Reserve Bank of India (RBI) regulations using Unsloth for efficient training. Achieved 57.6% accuracy (8.2x improvement over base model).

📊 Results

Metric	Base Model	Fine-tuned	Improvement
Overall Accuracy	7.0%	57.6%	+50.6%
Fact-based Questions	6.8%	57.6%	+50.8%
Reasoning Questions	37.5%	62.5%	+25.0%

8.2x better performance on RBI regulatory questions!

🎯 Project Overview

This project demonstrates end-to-end fine-tuning of a 3B parameter LLM for domain-specific question answering:

1. Data Collection: Scraping and processing RBI circulars
2. Dataset Creation: Generating 47K QA pairs with rephrasing for robustness
3. Fine-tuning: Using Unsloth for efficient LoRA training
4. Evaluation: Comprehensive testing with Gemini-based evaluation

Key Features

• ✅ Efficient Training: LoRA with Unsloth (2 hours on single GPU)
• ✅ Data Augmentation: 3 rephrased versions per question for generalization
• ✅ Production-Ready: Deployed model on Hugging Face
• ✅ Comprehensive Evaluation: 1000-sample stratified test set

🚀 Quick Start

Installation

pip install unsloth
pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
pip install --no-deps xformers trl peft accelerate bitsandbytes

Training

from unsloth import FastLanguageModel
from datasets import load_dataset

# Load model
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="Qwen/Qwen2.5-3B-Instruct",
    max_seq_length=2048,
    load_in_4bit=True,
)

# Apply LoRA
model = FastLanguageModel.get_peft_model(
    model,
    r=16,
    lora_alpha=32,
    lora_dropout=0.1,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
)

# Load dataset
dataset = load_dataset("Vishva007/RBI-Circular-QA-Dataset")

# Train (see training.ipynb for full code)

Inference

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "Vishva007/Qwen2.5-3B-Instruct-RBI-QA",
    max_seq_length=2048,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

messages = [
    {"role": "system", "content": "You are an expert on RBI regulations."},
    {"role": "user", "content": "What are the Basel III capital requirements?"}
]

inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")
outputs = model.generate(inputs, max_new_tokens=512, temperature=0.7, top_p=0.9)
print(tokenizer.decode(outputs, skip_special_tokens=True))

📁 Repository Structure

├── training.ipynb              # Complete training pipeline
├── eval.ipynb                  # Evaluation with before/after comparison
├── data_preparation/           # Scripts for dataset creation
│   ├── scrape_rbi.py          # RBI circular scraper
│   ├── generate_qa.py         # QA pair generation
│   └── rephrase_dataset.py    # Data augmentation via rephrasing
├── requirements.txt            # Python dependencies
└── README.md

🔧 Configuration

Training Hyperparameters

# LoRA Configuration
LORA_R = 16
LORA_ALPHA = 32
LORA_DROPOUT = 0.1

# Training
BATCH_SIZE = 8
GRADIENT_ACCUMULATION = 4  # Effective batch = 32
LEARNING_RATE = 2e-4
NUM_EPOCHS = 1
MAX_SEQ_LENGTH = 2048

# Hardware
GPU: NVIDIA L40S (44.5 GB VRAM)
Training Time: ~2 hours

Why This Configuration Works

• r=16: Optimal for 47K samples (not too small, not overfitting)
• alpha=32 (2×r): Stronger learning from rephrased data
• 1 epoch: Perfect for augmented dataset (4x exposure per concept)
• Cosine LR: Smooth convergence with warmup

See detailed explanation in training.ipynb.

📊 Dataset

RBI-Circular-QA-Dataset

• Size: 47,934 training samples + 1,000 eval samples
• Composition: 12K original + 35K rephrased QA pairs
• Coverage: 100+ regulation areas (Basel III, FEMA, AML, PSL, etc.)
• Time Range: RBI circulars from 2019-2024

Data Augmentation Strategy

Each original QA pair has 3 rephrased versions to teach conceptual understanding:

Original: "What relaxations were provided by RBI during COVID-19?"
Rephrase 1: "Can you describe the regulatory relief RBI offered during the pandemic?"
Rephrase 2: "How did RBI ease regulations in light of COVID-19?"
Rephrase 3: "Explain RBI's policy accommodations during the coronavirus crisis."

This prevents overfitting to exact phrasings and improves generalization.

📈 Evaluation

Comprehensive evaluation using Gemini 2.0 Flash as judge:

• Test Set: 1,000 stratified samples (balanced across regulation areas)
• Metrics: Binary pass/fail on factual accuracy
• Categories Tested: 100+ regulation areas, all institution types

Sample Results by Category

Category	Base	Fine-tuned	Δ
Anti-Money Laundering	5.4%	77.0%	+71.6%
Digital Payments	0.0%	77.8%	+77.8%
MSME Finance	12.5%	87.5%	+75.0%
Government Banking	0.0%	65.0%	+65.0%
Basel III Capital	4.5%	54.5%	+50.0%

See eval.ipynb for full evaluation pipeline.

📝 Technical Deep Dive

Want to understand the theory, methodology, and lessons behind this 8x improvement?

📖 Read the full article on DEV.to (10,000+ words)

Topics covered:

• 🧠 LoRA mathematics and configuration
• ⚡ Unsloth framework internals
• 📊 Data augmentation via rephrasing
• 🎓 Training hyperparameter analysis
• 📈 Evaluation methodology
• 🔬 Ablation studies and key lessons

Ideal for ML engineers building domain-specific LLMs or anyone wanting to replicate this approach.

🎓 Lessons Learned

What Worked

1. Data Augmentation via Rephrasing: Single biggest factor (8x improvement)
2. LoRA Efficiency: Only trained 1% of parameters, saved time and memory
3. Conservative Training: 1 epoch prevented overfitting on augmented data
4. Stratified Evaluation: Ensured reliable metrics across all categories

Key Insights

• Quality > Quantity: 47K well-augmented samples beat 100K+ generic samples
• Domain Specialization: Fine-tuned 3B model rivals RAG-enhanced GPT-4
• Evaluation Matters: Stratified sampling gave accurate performance estimates

🔮 Future Improvements

• Expand to 7B model for 65%+ accuracy
• Add retrieval (RAG) for 10-15% boost
• Preference optimization (DPO) on expert-labeled data
• Multi-lingual support (Hindi, regional languages)
• Real-time updates from new RBI circulars

🤝 Contributing

Contributions welcome! Areas for improvement:

• Additional regulation areas (SEBI, IRDAI, etc.)
• Better evaluation metrics
• Deployment optimizations (quantization, distillation)
• UI/API development

📄 License

This project is licensed under Apache 2.0.

🙏 Acknowledgments

• Unsloth: Fast and memory-efficient training
• Qwen Team: Excellent base model
• Hugging Face: Model hosting and datasets
• Google Gemini: Evaluation framework

📞 Contact

• Author: Vishva Ram
• HuggingFace: @Vishva007
• Model: Qwen2.5-3B-Instruct-RBI-QA
• Dataset: RBI-Circular-QA-Dataset

📚 Citation

@misc{vishva2025rbi-qa,
  author = {Vishva Ram},
  title = {Fine-tuning LLMs for RBI Regulatory Q&A},
  year = {2025},
  publisher = {GitHub},
  howpublished = {\url{https://github.com/vishvaRam/Unsloth-FineTuning}},
}

---

⭐ Star this repo if you found it useful!

Built with ❤️ for the Indian banking and AI community

This README includes:
- ✅ **Clear overview** with results upfront
- ✅ **Quick start** with code examples
- ✅ **Repository structure** for easy navigation
- ✅ **Detailed configuration** and explanations
- ✅ **Dataset information** with augmentation strategy
- ✅ **Evaluation methodology** and results
- ✅ **Lessons learned** (valuable for others)
- ✅ **Future roadmap** and contribution guidelines
- ✅ **Professional formatting** with badges and emojis