📨 Multi-Label Email Classifier using fine tuned LLaMA 3.2 (1B)

This project demonstrates how to build a multi-label email classification system using Meta LLaMa 3.2 (1B) model, fine-tuned using Parameter-Efficient Fine-Tuning (PEFT) and deployed using Gradio on Hugging Face Spaces. The model can be accessed with Multi_label Email Classifier

---

Overview

Given an email's subject and body, the model predicts one or more relevant categories from a predefined set.

---

Classification Categories

The model classifies emails into the following 10 categories:

CATEGORIES = [
    "Business", "Personal", "Promotions", "Customer Support", "Job Application",
    "Finance & Bills", "Events & Invitations", "Travel & Bookings",
    "Reminders", "Newsletters"
]

---

📊 Dataset Summary

We prepared a synthetic but realistic dataset of 2,105 labeled emails. Each email includes a subject, body, and one or more categories.

• Total Emails: 2105
• Label Distribution:

Business               : 941
Customer Support       : 227
Events & Invitations   : 656
Finance & Bills        : 334
Job Application        : 122
Newsletters            : 199
Personal               : 221
Promotions             : 120
Reminders              : 343
Travel & Bookings      : 314

The dataset can be accessed with Multi-label mail_dataset

---

Model Training

We fine-tuned meta-llama/Llama-3.1-1B-Instruct using the PEFT library with LoRA adapters. Training was done with:

• 🤗 Hugging Face transformers
• LoRA (for efficient fine-tuning)
• Resulting adapter uploaded to Hugging Face Hub:

imnim/multi-label-email-classifier Model card on hugging face

---

Backend & Deployment Options

Local (FastAPI + React)

We initially set up a FastAPI backend and a React frontend, but:

• The model was too large to run on CPU.
• No GPU available locally.
• Hugging Face Spaces does not support React + FastAPI apps with GPU on free tier.

Gradio on Hugging Face Spaces

Gradio apps run smoothly on Hugging Face's free CPU tier. Our Gradio version accepts subject and body inputs and returns predicted labels.

---

🌍 Access the App

➡️ https://huggingface.co/spaces/imnim/Multi-labelEmailClassifier

---