mlops-emotion-distilbert

Model Card: MLOps Emotion DistilBERT

Model Description

This is a fine-tuned version of distilbert-base-uncased for multi-class emotion classification. It was trained as part of an end-to-end MLOps pipeline assignment for the IIT Jodhpur PGD AI Program. The model takes English text as input and predicts one of six basic emotions.

• Base Model: distilbert-base-uncased
• Task: Text Classification
• Language: English
• License: IIT Jodhpur (All Rights Reserved)

Intended Uses & Limitations

This model is intended for educational purposes and basic emotion detection in short English text. It may not generalize well to complex, nuanced, or highly domain-specific language.

Training Data

The model was fine-tuned on the dair-ai/emotion dataset.

• Training Set: 16,000 samples
• Validation Set: 2,000 samples
• Classes:
  • 0: Sadness
  • 1: Joy
  • 2: Love
  • 3: Anger
  • 4: Fear
  • 5: Surprise

Training Procedure

The model was trained in a Kaggle Notebook utilizing dual T4 GPUs. Hyperparameter tuning was conducted to compare multiple versions, tracked via Weights & Biases.

The following hyperparameters were used for the primary version:

• Learning Rate: 3e-5
• Epochs: 2
• Train Batch Size: 16
• Eval Batch Size: 16
• Optimizer: AdamW

Evaluation Results

Model performance was evaluated on the validation split of the dataset. Metrics such as Accuracy and Weighted F1-score were monitored throughout training and logged directly to the project's W&B Dashboard.

Pipeline Integration

This model is containerized using Docker for inference and is integrated into a continuous CI/CD pipeline orchestrated via GitHub Actions.

• GitHub Repository: https://github.com/AzDevops143/MLOPSproj
• W&B Dashboard: https://wandb.ai/srajam696-charan/mlops-assignment3/overview