mjpsm / coca-cola-contact-classifier

Coca-Cola Contact Form Relevance Classifier

This repository contains a LoRA-fine-tuned text classification model designed to determine whether messages submitted through Coca-Cola’s contact form are business-relevant or not relevant .

The model is intended to support automated intake routing , helping prioritize legitimate sales, procurement, and partnership inquiries while filtering out consumer feedback, sponsorship requests, and unrelated messages.

🧠 Model Overview

• Base model: DistilBERT ( distilbert-base-uncased )
• Fine-tuning method: LoRA (Low-Rank Adaptation)
• Task: Binary text classification
• Labels:
  • relevant — business-related inquiries (sales, procurement, distribution, vendors)
  • not_relevant — consumer feedback, sponsorships, complaints, general messages
• Training size: 300 examples (balanced)
• Evaluation: Held-out test set (80/20 split)

🎯 Intended Use

This model is designed for:

• Contact form triage
• Business inquiry routing
• Intake prioritization workflows
• Manual review reduction

It is not intended to:

• Replace human judgment
• Make contractual or legal decisions
• Classify sentiment or emotions

📊 Performance Summary

Evaluated on a held-out test set:

Confusion Matrix: [28 1] [ 0 31]]

• The model prioritizes high recall , minimizing missed business inquiries.
• Borderline cases may be flagged for manual review using confidence thresholds.

⚠️ Important Notes on Confidence

The model intentionally outputs moderate confidence scores (typically 0.55–0.70) on ambiguous inputs.
This reflects realistic uncertainty and supports safe enterprise deployment.

Recommended usage:

• Confidence ≥ 0.70 → auto-route
• Confidence < 0.70 → manual review