roberta-base-go_emotions
| Property | Value |
|---|---|
| Parameter Count | 125M |
| License | MIT |
| Architecture | RoBERTa |
| Task Type | Multi-label Emotion Classification |
What is roberta-base-go_emotions?
roberta-base-go_emotions is a sophisticated emotion classification model built on the RoBERTa architecture, specifically trained to detect 28 different emotions in text. The model was trained on the go_emotions dataset, which consists of Reddit comments annotated with multiple emotion labels. This makes it particularly well-suited for understanding the complex emotional content in social media text.
Implementation Details
The model was trained using AutoModelForSequenceClassification with multi-label classification configuration. It underwent 3 epochs of training with a learning rate of 2e-5 and weight decay of 0.01. The model outputs 28 probability scores for each input text, typically using a 0.5 threshold for prediction.
- Multi-label classification architecture supporting 28 emotion categories
- Achieves F1 scores of 0.919 for gratitude and 0.829 for amusement
- Available in both PyTorch and ONNX formats (including INT8 quantized version)
- Optimized thresholds available for improved performance per emotion
Core Capabilities
- Emotion detection across 28 distinct categories
- Handles multiple emotions per text input
- Strong performance on common emotions like gratitude (F1: 0.919) and love (F1: 0.802)
- Specialized for social media content analysis
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its comprehensive coverage of 28 emotion categories and its ability to detect multiple emotions simultaneously in a single text. It's particularly notable for its strong performance on certain emotions like gratitude and amusement, making it ideal for social media sentiment analysis.
Q: What are the recommended use cases?
The model is best suited for analyzing social media content, customer feedback, or any text where multiple emotions might be present. It's particularly effective for detecting positive emotions like gratitude and amusement, though it may have limitations with less common emotions due to training data imbalance.
