Emotion Classification - ELECTRA-small (v2) - WINNING MODEL

This is the winning v2 model for our MLOps Group Project (Group 2, IIT Jodhpur).

Model Details

• Architecture: google/electra-small-discriminator
• Task: Text emotion classification (6 classes)
• Dataset: dair-ai/emotion
• Classes: sadness, joy, love, anger, fear, surprise

Training

• Platform: Kaggle GPU T4
• Epochs: 4
• Batch size: 16
• Learning rate: 5e-5
• Test Accuracy: 92.65%
• Test F1: 92.69%
• Model Size: 54.2 MB

Why This Won

Higher accuracy, lower loss, 2.5× smaller than v1 (MiniLM), and faster inference. The RTD pretraining objective produces more efficient representations.

Links

• GitHub Repo: https://github.com/nikhilsaini-iitj/MLOps_GroupProject
• Kaggle Notebook: https://www.kaggle.com/code/nikhilg25ait2067/kaggle-electra
• W&B Dashboard: https://wandb.ai/g25ait2067-prom-iit-rajasthan/mlops-groupproject-v2
• Docker Image: https://hub.docker.com/r/nikhilsainiiitj/mlops-groupproject-inference

Team

• Nikhil Saini (G25AIT2067)
• Y Sharathchandrika (G25AIT2132)
• Sarthak Kapoor (G25AIT2098)
• Aryaveer Rathi (G25AIT2021)