finmigodeveloper/gemma-3-270m-transaction-lora

🏦 Transaction Classification with Gemma-3 LoRA

📋 Model Overview

This is a LoRA adapter for google/gemma-3-270m-it fine-tuned on transaction categorization. It classifies financial transactions into 10 categories with 98.54% accuracy.

Categories

• Charity & Donations
• Entertainment & Recreation
• Financial Services
• Food & Dining
• Government & Legal
• Healthcare & Medical
• Income
• Shopping & Retail
• Transportation
• Utilities & Services

📊 Model Performance

Category	Accuracy
Charity & Donations	100.00%
Entertainment & Recreation	100.00%
Financial Services	100.00%
Food & Dining	98.45%
Government & Legal	97.40%
Healthcare & Medical	97.90%
Income	99.90%
Shopping & Retail	93.35%
Transportation	99.25%
Utilities & Services	98.80%
Overall	98.54%

🔧 LoRA Configuration

• Rank (r): 8
• Alpha: 32
• Dropout: 0.1
• Target Modules: q_proj, k_proj, v_proj, o_proj
• Trainable Parameters: 743,680 (0.28% of full model)
• Adapter Size: 2.98 MB

🚀 Quick Start

Installation

pip install torch transformers peft accelerate

Python Usage

from peft import PeftModel
from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

# Load base model
base_model = AutoModelForSequenceClassification.from_pretrained(
    "google/gemma-3-270m-it",
    num_labels=10,
    torch_dtype=torch.bfloat16
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "finmigodeveloper/gemma-3-270m-transaction-lora")
tokenizer = AutoTokenizer.from_pretrained("finmigodeveloper/gemma-3-270m-transaction-lora")

# Define categories
categories = [
    'Charity & Donations', 'Entertainment & Recreation', 'Financial Services',
    'Food & Dining', 'Government & Legal', 'Healthcare & Medical',
    'Income', 'Shopping & Retail', 'Transportation', 'Utilities & Services'
]
id2label = {i: label for i, label in enumerate(categories)}

# Classify a transaction
def classify_transaction(transaction_text):
    text = f"<start_of_turn>user\nClassify this transaction: {transaction_text}<end_of_turn>\n<start_of_turn>model\n"
    inputs = tokenizer(text, return_tensors="pt")
    outputs = model(**inputs)
    pred = torch.argmax(outputs.logits, dim=-1).item()
    return id2label[pred]

# Examples
print(classify_transaction("Starbucks coffee"))  # Food & Dining
print(classify_transaction("Uber ride"))  # Transportation
print(classify_transaction("Netflix subscription"))  # Entertainment & Recreation

Using with Hugging Face Inference API

import requests

API_URL = "/static-proxy?url=https%3A%2F%2Fapi-inference.huggingface.co%2Fmodels%2Ffinmigodeveloper%2Fgemma-3-270m-transaction-lora"
headers = {"Authorization": "Bearer YOUR_HF_TOKEN"}

def classify_transaction_api(text):
    payload = {"inputs": text}
    response = requests.post(API_URL, headers=headers, json=payload)
    return response.json()

# Example
result = classify_transaction_api("Starbucks coffee")
print(result)

🎯 Training Details

• Training Samples: 80,000 (10k per category)
• Validation Samples: 20,000
• Epochs: 3
• Batch Size: 32
• Gradient Accumulation: 2 steps
• Effective Batch Size: 64
• Learning Rate: 2e-4
• Optimizer: AdamW
• Training Time: 68 minutes on Tesla T4
• GPU Memory: 0.60 GB

🌐 Interactive Demo

Try the live demo on Hugging Face Spaces! (Create a Space first)

📈 Training Progress

Epoch	Training Loss	Validation Accuracy
1	0.0246	98.54%
2	0.0217	98.45%
3	0.0208	98.50%

⚠️ Limitations

• Shopping category (93.35% accuracy) could be improved with more diverse examples
• Best suited for English transactions
• May struggle with very short or ambiguous transaction descriptions
• Trained on transaction descriptions only (not full bank statements)

🔮 Future Improvements

• Add data augmentation for Shopping category
• Support multiple languages
• Add confidence thresholds for uncertain predictions
• Create ensemble with other models
• Add support for amount and currency context

📚 Citation

If you use this model, please cite:

@misc{gemma-3-270m-transaction-lora-2026,
  author = {finmigodeveloper},
  title = {Gemma-3 LoRA for Transaction Classification},
  year = {2026},
  publisher = {Hugging Face},
  journal = {Hugging Face Hub},
  howpublished = {\url{https://huggingface.co/finmigodeveloper/gemma-3-270m-transaction-lora}}
}

📄 License

This model is licensed under the Gemma license. See the original model page for details.

🤝 Contributing

Feel free to open issues or PRs if you have suggestions for improvements!

---

Made with ❤️ using Gemma-3 and LoRA