banking-intent-distilbert
distilbert-base-uncased fine-tuned for banking customer-intent classification over the
77 intents of the BANKING77 dataset (card issues, transfers, top-ups, fees, verification, …).
Training
- Method: full fine-tuning (all 66M parameters)
- Objective: cross-entropy, HF
Trainer - Data: mteb/banking77 — 10,003 train / 3,080 test, 77 classes
- Hyperparameters: 5 epochs, batch 32, lr 5e-5, max_len 64
- Hardware: CPU only (AMD Ryzen 9 7950X)
Results (test set)
| Epoch | Accuracy | Macro-F1 |
|---|---|---|
| 1 | 0.8072 | 0.7850 |
| 2 | 0.9060 | 0.9061 |
| 3 | 0.9226 | 0.9225 |
| 5 | 0.9278 | 0.9278 |
Macro-F1 tracks accuracy closely, indicating balanced performance across all 77 intents rather than only the frequent ones.
Usage
from transformers import pipeline
clf = pipeline("text-classification", model="vivekkopthsd/banking-intent-distilbert")
clf("My card payment was declined at the store")
# [{'label': 'declined_card_payment', 'score': 0.98}]
Predictions return readable intent names (e.g. declined_card_payment, card_arrival,
exchange_rate) via the model's id2label mapping.
Limitations
Trained on BANKING77, which is English-only, single-turn, and relatively clean. Real support traffic contains typos, code-switching, and multi-intent messages; expect degradation and re-tune on in-domain data before production use. Labels are BANKING77 class indices.