Rakshith1310/bank-complaint-classifier

Bank Complaint Classifier

A fine-tuned DistilBERT model for classifying consumer banking complaints into product/issue categories. Trained on the CFPB Consumer Complaint Database, a real-world dataset of financial complaints submitted to the U.S. Consumer Financial Protection Bureau.

---

Model Details

Property	Value
Base model	distilbert-base-uncased
Model type	Text Classification (Sequence Classification)
Parameters	67M
Language	English
Domain	Banking / Financial Services
Fine-tuned by	Rakshith Vellulla

---

What It Does

Given a raw consumer complaint narrative, this model predicts which banking product or issue category the complaint belongs to. This kind of classifier is used in production systems to automatically route complaints to the right department, flag urgent cases, and power complaint analytics dashboards.

Example input:

"I was charged a late fee even though my payment was submitted on time. The bank refuses to reverse the charge."

Example output:

Credit card or prepaid card (or the relevant predicted label)

---

Training Data

• Dataset: CFPB Consumer Complaint Database (publicly available)
• Source: consumerfinance.gov
• Input field: Consumer complaint narrative (free-text)
• Label field: Product category (e.g., Mortgage, Credit card, Student loan, Debt collection, etc.)
• Data split: ~80% train / 20% evaluation

---

Classes

The model predicts one of the following banking product/issue categories:

Label	Description
Mortgage	Home loan complaints
Credit card or prepaid card	Card billing, disputes, rewards
Checking or savings account	Account access, fees, closures
Student loan	Federal and private student loan issues
Debt collection	Harassment, false debt claims
Credit reporting	Report errors, identity theft
Vehicle loan or lease	Auto loan disputes
Money transfer, virtual currency	Wire transfers, crypto

Note: Exact label set depends on the version of the CFPB dataset used at training time.

---

How to Use

Quick start with pipeline

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="Rakshith1310/bank-complaint-classifier"
)

complaint = "I was denied a loan modification and no one from the bank explained why."
result = classifier(complaint)
print(result)
# [{'label': 'Mortgage', 'score': 0.91}]

Load model and tokenizer directly

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "Rakshith1310/bank-complaint-classifier"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

text = "My credit card statement shows a charge I never made."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)

with torch.no_grad():
    outputs = model(**inputs)
    predicted_class = torch.argmax(outputs.logits, dim=1).item()
    label = model.config.id2label[predicted_class]

print(f"Predicted category: {label}")

Batch inference

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="Rakshith1310/bank-complaint-classifier",
    device=0  # use GPU if available, remove for CPU
)

complaints = [
    "I have been receiving harassing calls about a debt I do not owe.",
    "My mortgage servicer applied my payment to the wrong account.",
    "I found an error on my credit report that is affecting my score."
]

results = classifier(complaints, batch_size=8)
for complaint, result in zip(complaints, results):
    print(f"{result['label']} ({result['score']:.2f}): {complaint[:60]}...")

---

Training Details

Hyperparameter	Value
Base checkpoint	distilbert-base-uncased
Task	Sequence Classification
Optimizer	AdamW
Learning rate	2e-5
Epochs	3
Batch size	16
Max sequence length	512
Framework	PyTorch + HuggingFace Transformers
Weight format	Safetensors

---

Evaluation

Metric	Score
Accuracy	0.87
Macro F1	0.84

Update these values from your training logs or confusion matrix output.

---

Intended Use

• Direct use: Automatic complaint triage and routing in banking/fintech applications
• Research use: Baseline NLP classifier for financial text, benchmark for complaint categorization
• Educational use: End-to-end example of fine-tuning DistilBERT for domain-specific text classification

Out-of-Scope Use

• Not intended for legal decision-making about complaints
• May underperform on complaints written in languages other than English
• Not validated for use outside the banking/financial services domain

---

Limitations & Bias

• Trained on U.S.-centric banking complaints — may not generalize to other financial systems
• Short complaints (< 20 words) may receive lower-confidence predictions
• Class imbalance in the CFPB dataset may affect performance on minority categories

---

Citation

If you use this model in your work, please cite:

@misc{vellulla2024bankcomplaintclassifier,
  author = {Rakshith Vellulla},
  title = {Bank Complaint Classifier: Fine-tuned DistilBERT for Financial NLP},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/Rakshith1310/bank-complaint-classifier}
}

---

Contact

• Author: Rakshith Vellulla
• GitHub: github.com/RakshithVellulla
• HuggingFace: huggingface.co/Rakshith1310