Ranjit0034/finee-llama-8b

FinEE Llama 8B - Indian Financial Entity Extractor

Model Description

FinEE Llama 8B is a fine-tuned version of Llama 3.1 8B Instruct, specialized for extracting financial entities from Indian banking messages (SMS, emails, statements).

Key Features

• 🏦 Multi-Bank Support: HDFC, ICICI, SBI, Axis, Kotak, and 20+ Indian banks
• 💳 All Transactions: UPI, NEFT, IMPS, Credit Card, EMI, Refunds
• 🌐 Multilingual: English, Hindi, Tamil, Telugu, Bengali, Kannada
• 📊 Structured Output: Clean JSON with all entities
• ⚡ Fast: <100ms per extraction (quantized)

Usage

With Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "Ranjit0034/finee-llama-8b",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Ranjit0034/finee-llama-8b")

message = "HDFC Bank: Rs.2,500 debited from A/c XX1234 on 12-Jan-26. UPI:swiggy@ybl. Ref:123456789012"

prompt = f"""Extract financial entities from this message:

{message}

JSON:"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
result = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(result)

With MLX (Apple Silicon)

from mlx_lm import load, generate

model, tokenizer = load("Ranjit0034/finee-llama-8b")
output = generate(model, tokenizer, prompt, max_tokens=256)
print(output)

With FinEE Package

from finee import FinancialExtractor

extractor = FinancialExtractor(model="Ranjit0034/finee-llama-8b")
result = extractor.extract("HDFC Bank: Rs.2,500 debited...")
print(result)
# {'amount': 2500.0, 'type': 'debit', 'merchant': 'Swiggy', 'category': 'food'}

Output Schema

{
  "amount": 2500.0,
  "type": "debit",
  "account": "1234",
  "bank": "HDFC",
  "date": "2026-01-12",
  "reference": "123456789012",
  "merchant": "Swiggy",
  "vpa": "swiggy@ybl",
  "category": "food",
  "is_p2m": true
}

Training

• Base Model: meta-llama/Llama-3.1-8B-Instruct
• Training Data: 152K+ samples (finee-dataset)
• Method: LoRA fine-tuning (rank=16)
• Hardware: Apple M2 Ultra (MLX)

Benchmarks

Metric	Score
Amount Accuracy	99.2%
Type Accuracy	98.5%
Merchant Detection	92.3%
Category Accuracy	88.7%
Overall F1	94.8%

Limitations

• Optimized for Indian banking messages
• May not work well with non-Indian formats
• Requires structured input (not handwritten)

Related

• 📦 FinEE Package - Python library
• 📊 Training Dataset
• 💻 GitHub

License

Apache 2.0