Ranjit0034/finance-entity-extractor

Finance Entity Extractor

A fine-tuned Phi-3 Mini model for extracting financial entities from transaction emails.

Model Description

This model extracts structured data from bank/payment notification emails:

• Amount (Rs. / ₹)
• Transaction type (debit/credit)
• Account number
• Date
• Reference number

Training

• Base model: Phi-3 Mini 4K Instruct
• Method: LoRA fine-tuning
• Framework: MLX (Apple Silicon optimized)
• Training data: 360 transaction emails
• Iterations: 500

Usage

from mlx_lm import load, generate

model, tokenizer = load("YOUR_USERNAME/finance-entity-extractor")

email = """
Dear Customer, Rs.2500.00 has been debited from account 3545 
to VPA swiggy@ybl on 28-12-25. Reference: 534567891234.
"""

prompt = f"Extract financial entities from this email:\n\n{email}"
response = generate(model, tokenizer, prompt=prompt, max_tokens=200)
print(response)

Output Format

{
  "amount": "2500",
  "type": "debit",
  "account": "3545",
  "date": "28-12-25",
  "reference": "534567891234"
}

Limitations

• Trained primarily on Indian bank emails (ICICI, HDFC)
• Works best with UPI/debit transactions
• Credit transactions may have lower accuracy

License

MIT