• Developed by: Team CodeBlooded
• Funded by: EpiUse & University of Pretoria
• Model type: DistilBertForSequenceClassification
• Language(s) (NLP): English

fin-classifier

Overview

Model Details

• Model type: DistilBertForSequenceClassification
• Version: v1.0 (initial release)
• Hugging Face repo: https://huggingface.co/CodeBlooded-capstone/fin-classifier
• Authors: CodeBlooded

Intended Use

Primary use case

• Task: Automated categorization of banking and credit card transaction descriptions for South Afrucan banks
• Users: Personal finance apps, budgeting tools, fintech platforms

Out-of-scope use cases

• Legal or compliance decisions
• Any use requiring 100% classification accuracy or safety guarantees

Training Data

• Source: Kaggle personal_transactions.csv dataset
• Mapping: Original vendor-level categories mapped into an internal schema of ~M high-level categories (data/categories.json).
• Feedback augmentation: User-corrected labels from feedback_corrected.json are appended to the training set for continuous improvement.

Evaluation

• Split: 90% train / 10% test split (seed=42) from the training corpus
• Metric: Macro F1-score
• Results:
  • Macro F1 on test set: 0.XX (not yet measured)

Usage

Limitations & Bias

• Performance varies by category: categories with fewer examples may see lower F1.
• The model reflects biases present in the original Kaggle dataset (e.g., over/under-representation of certain merchants).
• Should not be used as a sole source for financial decision-making.

Maintenance & Continuous Learning

• New user feedback corrections are stored in model/feedback_corrected.json and incorporated during retraining.
• Checkpoints are saved to results/ and versioned on Hugging Face.

License

Citation

@misc{fin-classifier2025,
  author = {CodeBlooded},
  title = {fin-classifier: A DistilBERT-based Transaction Categorization Model},
  year = {2025},
  howpublished = {\url{https://huggingface.co/CodeBlooded-capstone/fin-classifier}}
}