Dc-4nderson/email-importance

Email Importance Classification Dataset

Dataset Summary

This dataset is designed to train and evaluate text classification models on the task of distinguishing Important/Actionable emails from Noise/Promotional emails.

The data consists of raw email text bodies, simulating a realistic inbox environment containing personal correspondence, work requests, security alerts, newsletters, marketing blasts, and social media notifications. The text preserves realistic formatting, including newlines, messy URLs, and standard email headers, to ensure robustness in real-world applications.

Dataset Structure

Data Splits

Split	Rows	Description
Train	186	Used for training the model.
Validation	46	Used for evaluating model performance during training.
Total	232

Data Fields

• text (string): The raw content of the email, including subject lines (in some cases), salutations, body text, and footers.
• label_id (int): The classification label (0 or 1).

Label Legend

Label ID	Class Name	Description	Examples
0	Not Important	Low-priority, automated, or promotional content that does not require immediate action.	Marketing newsletters, social media "viewed profile" notifications, gamification updates (Duolingo), spam, receipts for low-value optional items.
1	Important	High-priority, personal, or transactional content that requires attention or action.	Job offers/interviews, direct personal emails, tax documents, security alerts (2FA, logins), bill overdue notices, shipping confirmations, flight cancellations.

Sample Data

{
  "text": "ACTION REQUIRED: W-2 Tax Form Available for 2025\n\nHi Dequan,\n\nYour 2025 W-2 tax form from TechSolutions Inc. is now available for download... ",
  "label_id": 1
},
{
  "text": "Huge Winter Blowout Sale! Up to 70% Off!\n\nCan't see images? Click here.\n\nLAST CHANCE TO SAVE...",
  "label_id": 0
}

Dataset Creation

The dataset was curated using a mix of synthetic generation and realistic templates to mimic the diversity of a modern inbox.

Personas: The dataset utilizes various recipient personas (e.g., Dequan, Elena, Marcus, Students, Professionals) to ensure the model does not overfit to a specific name or user profile.

Formatting: Great care was taken to include "messy" data, such as [Image] placeholders, long tracking URLs, unsubscribe footers, and HTML-to-text artifacts.

Usage

This dataset is ideal for:

Fine-tuning Small Language Models (SLMs) like DistilBERT, TinyLlama, or MobileBERT for edge-device email sorting.

Benchmarking zero-shot capabilities of larger LLMs on specific productivity tasks.

How to Load (Hugging Face Datasets)

from datasets import load_dataset

dataset = load_dataset("Dc-4nderson/email-importance")

print(dataset["train"][0])

Limitations

Synthetic PII: While the emails contain names, addresses, and phone numbers, all Personal Identifiable Information (PII) is synthetic/fictional. Any resemblance to real persons is coincidental.

Size: This is a small dataset (232 rows total) intended for few-shot learning, evaluation, or fine-tuning pre-trained models. It may not be sufficient for training a model from scratch.