merchant-consumption-category-discriminator-v1

Merchant Consumption Category Discriminator v1

• Repository: https://huggingface.co/kakao1513/merchant-consumption-category-discriminator-v1
• Base checkpoint: monologg/koelectra-base-v3-discriminator
• Export metadata model_name: monologg/koelectra-base-v3-discriminator
• Input format: merchant_text [SEP] normalized_merchant_text
• Validation macro F1: 0.6288
• Test macro F1: 0.6730
• Service fallback test macro F1: 0.6731
• Selected service fallback threshold: 0.3500

Summary

This model classifies Korean merchant strings into consumption categories for the more service. The labels are weakly supervised merchant-category labels derived from internal merchant-category pipelines, so boundary errors can still appear around visually similar service names.

Files

• Model weights and tokenizer are uploaded at the repository root.
• Training logs and evaluation artifacts are uploaded under artifacts/.

GroupKFold Results

• Number of folds: 5
• Best fold by validation macro F1: fold 3 (0.6763)
• Mean validation macro F1: 0.6532
• Std validation macro F1: 0.0231
• Mean validation accuracy: 0.6881
• Full fold metrics: https://huggingface.co/kakao1513/merchant-consumption-category-discriminator-v1/resolve/main/artifacts/kfold/cv_metrics.csv

Inference

from transformers import pipeline

clf = pipeline(
    'text-classification',
    model='kakao1513/merchant-consumption-category-discriminator-v1',
    tokenizer='kakao1513/merchant-consumption-category-discriminator-v1',
    top_k=3,
)

clf('스타벅스 강남R점 [SEP] 스타벅스 강남R점')

Limitations

• This classifier is optimized for merchant text classification, not full transaction understanding.
• Low-confidence predictions may still need the downstream service fallback rule.