mlops-emotion-classifier

Model Card for MLOps Emotion Classifier

Model Details

Model Description

This model is a fine-tuned version of distilbert-base-uncased trained on a small subset of the emotion dataset. It was developed specifically to fulfill the requirements of an End-to-End MLOps Pipeline academic assignment, demonstrating model fine-tuning, W&B tracking, Docker containerization, and GitHub Actions automation.

• Developed by: Group 15
• Funded by: Academic Assignment (PGD AI Program, IIT Jodhpur)
• Model type: Text Classification
• Language(s) (NLP): English
• License: MIT License
• Finetuned from model: distilbert-base-uncased

Model Sources

• Repository: https://github.com/g25ait2122/mlops-pipeline-project

Uses

Direct Use

This model is intended solely for educational purposes to demonstrate a working MLOps pipeline. It classifies English text into six basic emotions (sadness, joy, love, anger, fear, surprise).

Out-of-Scope Use

Because it was trained on a severely reduced dataset to meet Kaggle's free GPU time limits, this model is not suitable for any real-world or production applications.

Bias, Risks, and Limitations

The model inherits biases from the base DistilBERT model and the subset of the dair-ai/emotion dataset. Accuracy is intentionally compromised for the sake of pipeline speed.

How to Get Started with the Model

Use the code below to test the inference:

from transformers import pipeline

classifier = pipeline("text-classification", model="VaibhavG25AIT2122/mlops-emotion-classifier")
print(classifier("I am feeling very happy today!"))

Training Details

Training Data

Trained on a 2,000-sample subset of the dair-ai/emotion dataset to ensure fast training times on free compute.

Training Procedure

Preprocessing

Null values were dropped and all text was lowercased. Tokenized using DistilBertTokenizer with padding="max_length" and truncation=True.

Training Hyperparameters

• Training regime: fp32
• Learning rate: 5e-5
• Epochs: 2
• Batch size: 16

Evaluation

Testing Data, Factors & Metrics

Evaluated on a 500-sample validation subset using Accuracy and weighted F1-score.

Results

• Accuracy: ~82%
• F1 Score: ~0.81

Environmental Impact

• Hardware Type: GPU T4 x2
• Hours used: < 3 hours
• Cloud Provider: Kaggle Notebooks

Technical Specifications

Software

• transformers 4.38.2
• torch 2.2.1
• wandb
• datasets

Model Card Contact

Group 15 - Created for MLOps Assignment