Gandhiert/emotion-multilabel-distilbert

Emotion Multilabel Classification Model

🎭 A fine-tuned DistilBERT model for multilabel emotion classification on text data.

Model Description

This model is fine-tuned from distilbert-base-uncased for multilabel emotion classification. It can predict multiple emotions simultaneously from text input across 14 different emotion categories.

Performance

Kaggle Competition Results

• 🏆 Kaggle Test Score (Macro F1): 0.4214
• 📊 Validation Score (Macro F1): 0.4275
• 🎯 Generalization Gap: 0.0061 (excellent!)

Detailed Metrics

• Macro F1-Score: 0.4275 (validation) / 0.4214 (test)
• Micro F1-Score: 0.4226
• Hamming Loss: 0.1816
• Jaccard Score: 0.2679

Model Architecture

• Base Model: distilbert-base-uncased
• Parameters: 66,373,646 (~66M)
• Architecture: DistilBERT + Custom Classification Head
• Dropout: 0.3
• Loss Function: BCEWithLogitsLoss with class weights

Emotions Supported

The model can predict the following 14 emotions: ['amusement', 'anger', 'annoyance', 'caring', 'confusion', 'disappointment', 'disgust', 'embarrassment', 'excitement', 'fear', 'gratitude', 'joy', 'love', 'sadness']

Usage

from transformers import AutoTokenizer, AutoModel
import torch

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("your-username/emotion-multilabel-distilbert")
model = AutoModel.from_pretrained("your-username/emotion-multilabel-distilbert")

# Example usage
text = "I'm so happy and excited about this project!"
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=128)

with torch.no_grad():
    outputs = model(**inputs)
    predictions = torch.sigmoid(outputs.logits)
    
# Get emotions above threshold (0.5)
emotions = []
for i, prob in enumerate(predictions[0]):
    if prob > 0.5:
        emotions.append(emotion_columns[i])
        
print(f"Predicted emotions: {', '.join(emotions)}")

Training Details

• Training Duration: ~13 minutes
• Hardware: Tesla T4 GPU
• Epochs: 3
• Batch Size: 16
• Learning Rate: 2e-5
• Max Sequence Length: 128
• Optimizer: AdamW
• Class Weights: Applied for imbalanced dataset

Dataset Statistics

• Training Samples: 37,164
• Validation Samples: 9,291
• Test Samples: 8,199
• Average Labels per Sample: 3.21
• Most Common Pattern: 2-4 emotions per text

Performance Analysis

Strengths

• ✅ Good generalization (small validation-test gap)
• ✅ Reasonable multilabel predictions (avg 3.21 labels)
• ✅ Fast inference (~9 iterations/second)
• ✅ Memory efficient (66M parameters)

Areas for Improvement

• 📈 Macro F1 could be improved with hyperparameter tuning
• 📈 Class imbalance handling could be optimized
• 📈 Ensemble methods could boost performance

License

This model is released under the MIT License.