|
--- |
|
tags: |
|
- text-classification |
|
- multi-label |
|
- go-emotions |
|
- transformers |
|
- huggingface |
|
license: apache-2.0 |
|
library_name: transformers |
|
language: |
|
- en |
|
metrics: |
|
- accuracy |
|
- f1 |
|
base_model: |
|
- google-bert/bert-base-uncased |
|
pipeline_tag: text-classification |
|
--- |
|
|
|
# 🔥 Fine-Tuned BERT on GoEmotions Dataset |
|
|
|
## 📖 Model Overview |
|
This model is a **fine-tuned version of BERT** (`bert-base-uncased`) on the **GoEmotions** dataset for **multi-label emotion classification**. It can predict multiple emotions per input text. |
|
|
|
## 📊 Performance |
|
| Metric | Score | |
|
|----------------|-------| |
|
| **Accuracy** | 46.57% | |
|
| **F1 Score** | 56.41% | |
|
| **Hamming Loss** | 3.39% | |
|
|
|
## 📂 Model Usage |
|
```python |
|
from transformers import AutoTokenizer, AutoModelForSequenceClassification |
|
import torch |
|
|
|
# Load model and tokenizer |
|
model_name = "codewithdark/bert-Gomotions" |
|
tokenizer = AutoTokenizer.from_pretrained(model_name) |
|
model = AutoModelForSequenceClassification.from_pretrained(model_name) |
|
|
|
# Emotion labels (adjust based on your dataset) |
|
emotion_labels = [ |
|
"Admiration", "Amusement", "Anger", "Annoyance", "Approval", "Caring", "Confusion", |
|
"Curiosity", "Desire", "Disappointment", "Disapproval", "Disgust", "Embarrassment", |
|
"Excitement", "Fear", "Gratitude", "Grief", "Joy", "Love", "Nervousness", "Optimism", |
|
"Pride", "Realization", "Relief", "Remorse", "Sadness", "Surprise", "Neutral" |
|
] |
|
|
|
# Example text |
|
text = "I'm so happy today!" |
|
inputs = tokenizer(text, return_tensors="pt") |
|
|
|
# Predict |
|
with torch.no_grad(): |
|
outputs = model(**inputs) |
|
probs = torch.sigmoid(outputs.logits).squeeze(0) # Convert logits to probabilities |
|
|
|
# Get top 5 predictions |
|
top5_indices = torch.argsort(probs, descending=True)[:5] # Get indices of top 5 labels |
|
top5_labels = [emotion_labels[i] for i in top5_indices] |
|
top5_probs = [probs[i].item() for i in top5_indices] |
|
|
|
# Print results |
|
print("Top 5 Predicted Emotions:") |
|
for label, prob in zip(top5_labels, top5_probs): |
|
print(f"{label}: {prob:.4f}") |
|
|
|
''' |
|
output: |
|
Top 5 Predicted Emotions: |
|
Joy: 0.9478 |
|
Love: 0.7854 |
|
Optimism: 0.6342 |
|
Admiration: 0.5678 |
|
Excitement: 0.5231 |
|
''' |
|
``` |
|
|
|
## 🏋️♂️ Training Details |
|
- **Model:** `bert-base-uncased` |
|
- **Dataset:** [GoEmotions](https://huggingface.co/datasets/go_emotions) |
|
- **Optimizer:** AdamW |
|
- **Loss Function:** BCEWithLogitsLoss (Binary Cross-Entropy for multi-label classification) |
|
- **Batch Size:** 16 |
|
- **Epochs:** 3 |
|
- **Evaluation Metrics:** Accuracy, F1 Score, Hamming Loss |
|
|
|
## 📌 How to Use in Hugging Face |
|
```python |
|
from transformers import pipeline |
|
|
|
classifier = pipeline("text-classification", model="codewithdark/bert-Gomotions", top_k=None) |
|
classifier("I'm so excited about the trip!") |
|
``` |
|
|
|
## 🛠️ Citation |
|
If you use this model, please cite: |
|
```bibtex |
|
@misc{your_model, |
|
author = {codewithdark}, |
|
title = {Fine-tuned BERT on GoEmotions}, |
|
year = {2025}, |
|
url = {https://huggingface.co/codewithdark/bert-Gomotions} |
|
} |
|
``` |
