Base model: answerdotai/ModernBERT-base (149M params, Apache-2.0)
Task: single-label intent detection over 77 fine-grained banking intents
Held-out test: the official Banking77 test split (3,080 queries, 40/intent), disjoint from training
Result: 0.9399 accuracy / 0.9401 macro-F1 on the test split
Training code: https://github.com/sukhrobnurali/banking77-modernbert

Intended use

Routing short English banking/customer-support queries to one of 77 intents (e.g. card_arrival, lost_or_stolen_card, exchange_rate) for task-oriented dialog and triage.

Out of scope

Non-English text and non-banking domains (single-domain training data).
Out-of-scope / rejection detection: every input is forced into one of 77 intents.
Long documents — trained at max_length=64 on short single-sentence queries.

Evaluation

The same protocol is applied to every row of the table, all on the untouched test split. Two baselines contextualize the fine-tune's gain: a majority-class floor, and a frozen-encoder linear probe (mean-pooled base-model embeddings → logistic regression).

Model	Accuracy	Macro-F1	Weighted-F1
Majority class	0.0130	0.0003	0.0003
Frozen ModernBERT + linear probe	0.8945	0.8948	0.8948
This model (fine-tuned)	0.9399	0.9401	0.9401

ModernBERT's pretrained representations are already strong on this task — a linear probe on frozen embeddings reaches 0.8948 macro-F1 — so fine-tuning adds +4.5 points of macro-F1 (0.8948 → 0.9401) on top of that, well clear of the 0.0003 majority-class floor.

At ~0.94 accuracy the residual errors stay concentrated among semantically adjacent intents — the closely related transfer / top-up families and look-alike card intents. Full per-class breakdown: results/classification_report.txt; 77×77 confusion matrix: results/confusion_matrix.png.

Usage

python

from transformers import pipeline

clf = pipeline("text-classification", model="sukhrobnurali/modernbert-base-banking77")
clf("My card still hasn't arrived, when will I get it?")
# -> [{'label': 'card_arrival', 'score': 0.99}]

Training data

PolyAI/banking77 (Casanueva et al. 2020, arXiv:2003.04807): 10,003 train / 3,080 test online-banking queries over 77 intents. License CC-BY-4.0. A stratified 10% of train was held out as a validation set for early stopping; the official test split was used only for the final numbers above.

Reproducibility

Fixed seed=42. Fine-tuned with the HF Trainer, lr=5e-5, up to 4 epochs with early stopping (patience 2) on validation macro-F1, max_length=64, batch size 32, warmup ratio 0.1, weight decay 0.01. Pinned versions in requirements.txt (transformers 5.10.1). eval.py reproduces the test numbers from the published model.

License

Apache-2.0, inherited from the base model answerdotai/ModernBERT-base. Training data PolyAI/banking77 is CC-BY-4.0 (attribution: Casanueva et al. 2020).

Citation

bibtex

@misc{nurali_modernbert_banking77_2026,
  author       = {Sukhrob Nurali},
  title        = {modernbert-base-banking77: a ModernBERT intent classifier for Banking77},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/sukhrobnurali/modernbert-base-banking77}}
}

Author

Sukhrob Nurali
Hugging Face: sukhrobnurali
GitHub: sukhrobnurali