indian-txn-classifier

Indian Transaction Classifier

Fine-tuned Qwen2.5-0.5B for classifying Indian bank transactions (UPI, NEFT, IMPS, RTGS) into 30+ categories with merchant identification.

⚡ Try it now (Inference API — free, no setup)

Use the Inference API widget on the right side of this page. Type a transaction description in the text box and click Compute.

Or call it programmatically:

from huggingface_hub import InferenceClient

client = InferenceClient(model="SahilGoel/indian-txn-classifier")

system_prompt = 'You are a bank transaction classifier for Indian bank statements. Given a raw transaction description, infer both its category and the actual company when evidence exists. Respond with ONLY a JSON object: {"category": "<category>", "company_name": "<company_or_null>", "is_income": false, "confidence": 0.0}.'

tx = "UPI/zerodhabroking@/HDFC BANK LTD"
prompt = f"### System:\n{system_prompt}\n\n### Input:\n{tx}\n\n### Output:\n"

result = client.text_generation(prompt, max_new_tokens=100, temperature=0.1)
print(result)
# {"category": "trading_deposit", "company_name": "Zerodha", "is_income": false, "confidence": 0.95}

Categories

Income: salary, dividend, interest, rental, capital_gains, other_income

Expenses: food, grocery, shopping, bills, medical, insurance, tax_payment, credit_card, personal_transfer, investment, trading_deposit, trading_credit, education, travel, entertainment, donation, loan_emi, loan_repayment, cash_withdrawal

Special: friends, family, flat_deposit, trading_fees, vehicle_purchase, staff_salary, health_fitness, transfer, unclassified

Pipeline Architecture

1. Rule engine (70-80% coverage) — regex patterns for known merchants and UPI handles
2. Recurring pattern detector (10% more) — identifies repeating transactions
3. Fine-tuned Qwen 0.5B (remaining) — LLM fallback for uncertain transactions

Run locally

pip install transformers torch

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("SahilGoel/indian-txn-classifier")
tokenizer = AutoTokenizer.from_pretrained("SahilGoel/indian-txn-classifier")

system_prompt = "You are a bank transaction classifier for Indian bank statements..."
input_text = "UPI/swiggybengaluru@/HDFC BANK LTD"
prompt = f"### System:\n{system_prompt}\n\n### Input:\n{input_text}\n\n### Output:\n"

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100, temperature=0.1, do_sample=False)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Run on Google Colab (free GPU)

Open app.py from the GitHub repo in a Colab notebook — it works as a Gradio app with a public share link.

Supported Banks

ICICI, HDFC, SBI, Axis, Kotak, Yes Bank, Federal Bank, IDFC First, IndusInd, Bank of Baroda, Punjab National, Canara, Union Bank, Unity SFB

GitHub

Code and training pipeline: the-great-one/indian-txn-classifier