sg-sentiment-roberta

Singapore Sentiment Analyzer - ROBERTA (Calibrated)

Fine-tuned sentiment analysis model for Singapore social media, with post-training calibration for improved accuracy.

🎯 Performance

📊 Sentiment Scale

🚀 Quick Start

from transformers import AutoTokenizer
from modeling_calibrated import CalibratedRegressionModel

# Load model (calibration is automatic!)
model_name = "your-username/roberta-singapore-sentiment"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = CalibratedRegressionModel.from_pretrained(model_name)

# Predict sentiment
text = "This chicken rice is damn shiok sia!"
result = model.predict_sentiment(text, tokenizer)

print(f"Score: {result['score']:.3f}")      # 0.875
print(f"Category: {result['category']}")   # "Very Positive"

💡 What is Calibration?

After fine-tuning, we applied isotonic regression calibration on a validation set. This corrects systematic bias patterns where the model was:

• Over-predicting on negative examples
• Under-predicting on positive examples
• Struggling with boundary cases (e.g., neutral vs negative)

The calibration layer is built into the model - you get calibrated predictions automatically!

📚 Training Details

• Base model: cardiffnlp/twitter-roberta-base-sentiment-latest
• Training data: 49,521 Singapore Reddit posts/comments
• Fine-tuning: 5 epochs, MSE loss, learning rate 2e-5
• Calibration: Isotonic regression on 500-sample validation set

🌏 Singapore Context

This model understands Singlish patterns and Singapore-specific terminology:

• Particles: lah, lor, leh, sia
• Slang: shiok, sian, jialat, paiseh
• Local context: HDB, MRT, hawker, kopitiam

📝 Citation

@misc{roberta-singapore-calibrated,
  title = {Singapore Sentiment Analyzer - ROBERTA (Calibrated)},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/your-username/roberta-singapore-sentiment}
}

📄 License

MIT License - Free for commercial and non-commercial use.