Receipt-Classifier

Project: Makerspace Inventory Management Project Author: Ethan Kessler (Carnegie Mellon University) License: MIT Date: 2025

Description:

This model predicts whether a line from a receipt received from McMaster Carr contains important information for describing item name, quantity, vendor, manufacturer, or other important inventory field listing. It was trained using using code generated by Copilot on a curated dataset with line information extracted using Tesseract OCR from 20 receipts retrieved from https://www.mcmaster.com/.

Framework:

• Model: WeightedDistilbert
• Cross Entropy Loss: 'Other' = 1.0,'Imporant' = 10.0
• Epochs: 6
• batch size per device: 16
• Learning rate: 5e-5
• Weight decay: 0.01

Performance:

• eval_loss ≈ 0.06072889268398285
• eval_accuracy ≈ 0.9861111111111112
• eval_f1 ≈ 0.9411764705882353
• eval_runtime: 2.1135
• eval_samples_per_second: 68.132
• eval_steps_per_second: 8.517

Notes:

• Intended for referee-assistive scoring and highlight extraction.
• Trained on clean data for Olympic-level scenarios. Limitations:

---

• Trained only on receipts from McMaster-Carr.
• May not identify prices.

Ethical Use:

For research, education, and makerspace management only. All data sourced from McMaster-Carr. Citation:

Kessler, E. (2025). "receipt-classifier" Hugging Face: https://huggingface.co/emkessle/receipt-classifier