Fine-tuned answerdotai/ModernBERT-base on PolyAI/banking77 (77 fine-grained banking intents). Part of Sift, a query-triage project showing a small fine-tuned model beats a zero-shot open 7B instruct LLM (Mistral-7B-Instruct, run locally and 4-bit quantised) on cost, latency, and privacy for narrow, high-volume classification.

Metrics — full test split (3076 examples)

Metric	Score
Accuracy	0.9353
Macro-F1	0.9353
Median single-query latency	22.47 ms (Tesla T4)

Intent -> team routing

The 77 intents are mapped to six routing teams: Cards, Payments, Accounts, Fraud/Security, Onboarding, General.

Usage

python

from transformers import pipeline
clf = pipeline("text-classification", model="DarioFrK/sift-banking77-modernbert", top_k=5)
clf("My card was swallowed by the ATM")

Base model: answerdotai/ModernBERT-base
Dataset: PolyAI/banking77
Training: 3 epochs, lr=5e-05, batch=16, max_len=64