belfidi/distilbert-intent-bitext

DistilBERT intent classifier — Bitext customer support (6 classes)

Fine-tuned distilbert-base-uncased for customer-service intent recognition, built for a university project (IU International University of Applied Sciences, DLBAIBEPNBACSA01).

Labels

The model config uses generic label names. Index → intent (alphabetical LabelEncoder order):

Index	Config label	Intent
0	LABEL_0	account_management
1	LABEL_1	billing_payment
2	LABEL_2	complaint_service
3	LABEL_3	order_tracking
4	LABEL_4	product_inquiry
5	LABEL_5	refund_cancellation

The 27 original Bitext intents were mapped to these 6 categories.

Training

• Data: Bitext Customer Support (26,872 utterances); URLs → <URL>, mentions → <USER>
• Split: stratified 80/20, random_state=42
• 3 epochs, learning rate 2e-5, batch size 16, max_length=64, seed 42

Evaluation (validation set, n = 5,375)

Model	Accuracy	Errors	Macro-F1
This model	99.96 %	2	0.9995
TF-IDF + Logistic Regression (baseline)	99.74 %	14	0.9974

Bitext utterances are synthetic and highly regular, so these scores overstate performance on real customer messages. Out-of-scope queries can be assigned to one of the six intents with high confidence.