Fine-tuned sentence-transformers/all-MiniLM-L6-v2 for 23-class transaction categorization. Trained on the Finehance synthetic transaction corpus (~110k deduped samples covering processor prefixes, store numbers, OCR-style truncations, multilingual descriptors).

Held-out eval

Accuracy: 0.9656
F1 (macro): 0.9663
F1 (weighted): 0.9656

Labels

Bills & Utilities
Cash & ATM
Childcare
Coffee & Beverages
Convenience
Education
Entertainment
Fast Food
Food Delivery
Gas & Fuel
Giving
Groceries
Healthcare
Housing
Income
Insurance
Other
Restaurants
Shopping & Retail
Subscriptions
Transfers
Transportation
Travel

Usage

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tok = AutoTokenizer.from_pretrained("CyberKunju/finehance-categorizer-minilm")
model = AutoModelForSequenceClassification.from_pretrained("CyberKunju/finehance-categorizer-minilm")
model.eval()

desc = "UBER *EATS HELP.UBER.COM"
inputs = tok(desc, return_tensors="pt", truncation=True, max_length=64)
with torch.no_grad():
    logits = model(**inputs).logits
pred_id = logits.argmax(-1).item()
print(model.config.id2label[pred_id])

Intended as the MiniLM arm of a 3-way probability-blend ensemble (TF-IDF LR ⊕ LightGBM on MiniLM embeddings ⊕ this classifier) inside the Finehance Tier 2 categorizer.