Goemotions Bert

🚀 GoEmotions BERT Classifier

Fine-tuned BERT model for multi-label emotion classification.

This project presents a fine-tuned BERT-base-uncased model on the go_emotions dataset for multi-label classification of 28 emotions.

✨ Features

• Multi-label Classification: Capable of classifying text into 28 different emotions.
• Optimized Thresholds: Ensures more accurate predictions.
• High Performance: Achieves good scores on various evaluation metrics.

📦 Installation

No specific installation steps are provided in the original document.

🚀 Quick Start

For accurate predictions with optimized thresholds, use the Gradio demo.

📚 Documentation

Model Details

Performance

• Micro F1: 0.6025 (optimized thresholds)
• Macro F1: 0.5266
• Precision: 0.5425
• Recall: 0.6775
• Hamming Loss: 0.0372
• Avg Positive Predictions: 1.4564

Class-Wise Performance

The following table shows per-class metrics on the test set using optimized thresholds (see thresholds.json):

💻 Usage Examples

Basic Usage

from transformers import BertForSequenceClassification, BertTokenizer
import torch
import json
import requests

# Load model and tokenizer
repo_id = "logasanjeev/goemotions-bert"
model = BertForSequenceClassification.from_pretrained(repo_id)
tokenizer = BertTokenizer.from_pretrained(repo_id)

# Load thresholds
thresholds_url = f"https://huggingface.co/{repo_id}/raw/main/thresholds.json"
thresholds_data = json.loads(requests.get(thresholds_url).text)
emotion_labels = thresholds_data["emotion_labels"]
thresholds = thresholds_data["thresholds"]

# Predict
text = "I’m just chilling today."
encodings = tokenizer(text, padding='max_length', truncation=True, max_length=128, return_tensors='pt')
with torch.no_grad():
    logits = torch.sigmoid(model(**encodings).logits).numpy()[0]
predictions = [(emotion_labels[i], logit) for i, (logit, thresh) in enumerate(zip(logits, thresholds)) if logit >= thresh]
print(sorted(predictions, key=lambda x: x[1], reverse=True))
# Output: [('neutral', 0.8147)]

📄 License

This project is licensed under the MIT license.