Ahmed-Tarek0/complaint-classifier

Consumer Complaint Classifier

A fine-tuned DistilBERT model that classifies a consumer complaint narrative into one of five categories. Built as part of an end-to-end NLP project comparing SimpleRNN, LSTM, GRU, and a fine-tuned transformer on the same data.

Labels

Label	Description
credit_card	Complaints about credit card charges, fraud, fees
credit_reporting	Complaints about credit reports and credit bureaus
debt_collection	Complaints about debt collection practices
mortgages_and_loans	Complaints about mortgages and loan applications
retail_banking	Complaints about checking/savings accounts and retail banking

Performance

Evaluated on a held-out test set, compared against three models trained from scratch on the same split:

Model	Accuracy	Precision	Recall	F1-score
SimpleRNN	0.749	0.784	0.749	0.759
LSTM	0.842	0.858	0.842	0.846
GRU	0.850	0.861	0.850	0.853
This model (fine-tuned DistilBERT)	0.883	0.887	0.883	0.884

How to use

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="Ahmed-Tarek0/complaint-classifier"
)

classifier("I was charged twice for the same transaction and the bank refuses to refund me.")

Or load the tokenizer and model directly:

from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("Ahmed-Tarek0/complaint-classifier")
model = AutoModelForSequenceClassification.from_pretrained("Ahmed-Tarek0/complaint-classifier")

Training details

• Base model: distilbert-base-uncased
• Task: Multi-class text classification (5 classes)
• Fine-tuning: Hugging Face Trainer, weighted cross-entropy loss to handle class imbalance, early stopping on validation F1
• Max sequence length: 150 tokens
• Trained alongside SimpleRNN, LSTM, and GRU baselines built from scratch for comparison — see the project notebook for the full pipeline.

Intended use

Built for a classification/deployment learning project. Suitable for demonstrating complaint-routing use cases; not validated for production financial compliance use without further evaluation.