MithunKaveriappa/banking77-distilbert

Banking77 Intent Classifier (DistilBERT)

DistilBERT fine-tuned to classify customer messages into 77 banking intents (e.g. card_arrival, top_up_failed, pending_transfer).

Results (test set, 3,076 examples)

Model	Accuracy	Macro-F1
TF-IDF + Logistic Regression	0.8898	0.8899
BGE-small embeddings + Logistic Regression	0.9311	0.9310
DistilBERT (this model)	0.9217	0.9217

Training

• Base model: distilbert-base-uncased
• Data: mteb/banking77, 90/10 train/validation split (stratified), test set held out
• 5 epochs, learning rate 5e-5, batch size 32, max length 64, fp16 on a T4 GPU
• Best epoch selected by validation macro-F1

Usage

from transformers import pipeline
clf = pipeline("text-classification", model="MithunKaveriappa/banking77-distilbert", top_k=3)
print(clf("My card has not arrived yet"))

Limitations

• Trained only on short English banking messages.
• Always predicts one of the 77 known intents, even for unrelated text, so low confidence scores should be treated as "unknown".
• Several intents overlap (e.g. top-up failed vs. reverted), which causes most errors.