fahadkamraan/transaction-categorizer

Transaction Categoriser

Fine-tuned DistilBERT model that classifies raw bank transaction descriptions into 17 merchant spending categories with 99.88% test accuracy.

Try it live: 🤗 Spaces Demo

---

Model Details

Property	Value
Base model	distilbert-base-uncased
Task	Multi-class text classification (17 classes)
Parameters	66M
Training platform	Kaggle (GPU T4 x2)
Framework	HuggingFace Transformers 4.44.2
Inference	CPU-compatible

Labels

{
  0: "Education",       1: "Entertainment",   2: "Fees",
  3: "Groceries",       4: "Healthcare",       5: "Income",
  6: "Insurance",       7: "Mortgage",         8: "Personal Care",
  9: "Rent",           10: "Restaurants",     11: "Shopping",
 12: "Subscription",   13: "Transfer",        14: "Transportation",
 15: "Travel",         16: "Utilities"
}

Performance

Evaluated on a held-out test set of 4,298 samples (10% of 42,975 total).

Metric	V1 (baseline)	V2 (this model)
Accuracy	99.70%	99.88%
F1 (weighted)	99.70%	99.88%
F1 (macro)	99.71%	99.89%

All experiments tracked on W&B.

Usage

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="fahadkamraan/transaction-categorizer",
    device=-1,   # CPU
    top_k=3,
)

results = classifier("starbucks coffee purchase")
# [{'label': 'Restaurants', 'score': 0.9999}]

With HuggingFace Hub

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("fahadkamraan/transaction-categorizer")
model = AutoModelForSequenceClassification.from_pretrained("fahadkamraan/transaction-categorizer")

inputs = tokenizer("netflix monthly subscription", return_tensors="pt", truncation=True, max_length=64)
with torch.no_grad():
    logits = model(**inputs).logits
predicted_class = logits.argmax(-1).item()
print(model.config.id2label[predicted_class])  # → Subscription

Docker

docker run --rm \
  -e HF_TOKEN=<token> \
  -e INPUT_TEXT="delta airlines flight booking" \
  fahadkamraan/mlops-transaction-classifier:latest

Training Data

• Dataset: DoDataThings/us-bank-transaction-categories-v2
• Licence: MIT
• Raw size: 68,000 rows · 17 categories
• Used: 42,975 rows (up to 3,000 per class after deduplication)
• Split: 80% train / 10% val / 10% test

Training Procedure

Two versions were trained and compared on Kaggle (GPU T4 x2):

Hyperparameter	V1	V2 (this model)
Learning rate	2e-5	5e-5
Epochs	3	5
Weight decay	0.0	0.01
Preprocessing	Light	Extra normalisation
Batch size	32	32

V2 preprocessing additionally strips [debit]/[credit] prefix tags and standardises store number suffixes, giving the model cleaner input.

Limitations

• Supports only the 17 categories present in the training dataset. Transactions from categories like "Investments" or "Taxes" will be mapped to the nearest available label.
• Trained on US-style bank transaction descriptions. May generalise poorly to non-English or very different regional bank formats.
• Maximum input length: 64 tokens (sufficient for typical transaction strings).

MLOps Pipeline

This model is part of a complete MLOps assignment pipeline:

• GitHub: S-FK/transaction_categorization
• Docker: fahadkamraan/mlops-transaction-classifier
• W&B: mlops-transaction-classifier
• Spaces Demo: transaction-categorizer-demo

Citation

@misc{kamraan2025transactioncategorizer,
  author    = {S Fahad Kamraan and Dhruvi Patel and Mahesh V and Himanshu Choubey},
  title     = {Bank Transaction Merchant Categoriser},
  year      = {2025},
  publisher = {HuggingFace Hub},
  url       = {https://huggingface.co/fahadkamraan/transaction-categorizer}
}

IIT Jodhpur — ML Ops Assignment (CSL 7040)