1. **Spam vs. Non‑spam classification** – The model must label each email as “spam” or “ham” based on content, sender reputation, and header features, with a high penalty for misclassifying spam as ham.

2. **Urgency detection** – Emails are scored on a 0–1 urgency scale, where the system must predict whether an email requires an action within 2 hours, 24 hours, or can wait. The benchmark includes a mix of time‑sensitive notifications, meeting invites, and routine updates.

3. **Attachment risk assessment** – For each email with one or more attachments, the model must predict a risk score (0–1) indicating the likelihood that the attachment is malicious. The data set contains benign PDFs, images, executables, and a few disguised malicious files.

4. **Follow‑up necessity prediction** – The system determines whether an email needs a reply or a follow‑up action. The benchmark contains responses to customer support tickets, internal project updates, and social messages, with a balanced mix of “needs reply” and “no action required”.

5. **Categorization by topic** – Emails are assigned to one of five broad categories (e.g., “Finance”, “HR”, “Engineering”, “Marketing”, “Personal”). The benchmark includes synthetic subject lines and body text that mimic realistic business and personal email patterns.