DistilBERT for Emotion Classification (IIT Jodhpur MLOps Group 14)
This model is a fine-tuned version of distilbert-base-uncased trained on the dair-ai/emotion dataset. It was developed as part of an end-to-end MLOps pipeline assignment to classify text into six core emotions: sadness, joy, love, anger, fear, and surprise.
📊 Experimentation Summary
We conducted multiple tracking experiments logged via Weights & Biases to discover the optimal hyperparameters:
| Experiment Version | Learning Rate | Epochs | Weight Decay | Eval Accuracy | Status |
|---|
| Experiment 1 (Baseline) | 2e-5 | 3 | 0.01 | 90.4% | Deprecated |
| Experiment 2 (Tuned) | 5e-5 | 4 | 0.05 | 93.7% | Production Candidate |
🛠️ Training Hyperparameters
The winning model (this checkpoint) was trained with the following configurations:
- Learning Rate: 5e-5
- Train Batch Size: 16
- Eval Batch Size: 16
- Epochs: 4
- Weight Decay: 0.05
- Optimizer: AdamW
📈 Evaluation Results
The model achieved a final validation accuracy of 93.7%, demonstrating robust convergence and strong generalization across all 6 emotional categories.
🧪 Tracked with Weights & Biases
All training runs, real-time gradients, and validation loops were monitored transparently. You can inspect our charts under the project workspace: mlops-emotion-classification.
