cinema-verde-spam-classifier

Cinema Verde Spam Classifier

A fine-tuned DistilBERT model for binary email classification, trained to distinguish legitimate internship inquiries from B2B marketing spam for Cinema Verde, an environmental film festival.

Model Details

• Base model: distilbert-base-uncased
• Task: Binary text classification (LEGIT / SPAM)
• Framework: HuggingFace Transformers + PyTorch

Training Data

• 150 real labeled emails collected during internship operations
• Dataset had a 2:1 class imbalance (100 legit, 50 spam)
• Resolved via synthetic data augmentation using the Gemini API to generate 50 additional spam examples, producing a balanced 200-email training set

Results

The synthetic augmentation model was selected for deployment due to its balanced precision and recall, reducing false positives on legitimate internship applications.

Intended Use

Deployed as part of a Gmail Add-on (InternSift) to help small organizations filter internship inquiry emails from B2B spam in real time.

Limitations

• Trained on a small dataset (200 examples); performance may vary on unseen spam patterns
• Optimized specifically for Cinema Verde's email context