Multilabel Emotion Classification Model - FirstTimeUp

This model is fine-tuned for multilabel emotion classification using distilbert-base-uncased as the base model.

Model Details

Model Name: FirstTimeUp
Base Model: distilbert-base-uncased
Task: Multilabel Emotion Classification
Emotions: amusement, anger, annoyance, caring, confusion, disappointment, disgust, embarrassment, excitement, fear, gratitude, joy, love, sadness
Total Parameters: 66,373,646
Trainable Parameters: 66,373,646

Quick Start

Installation

pip install torch transformers huggingface_hub

Usage

# Download the repository
from huggingface_hub import snapshot_download
import sys
import os

# Download model files
model_path = snapshot_download(repo_id="EnJiZ/FirstTimeUp")

# Add to path and import
sys.path.append(model_path)
from model import predict_emotions

# Predict emotions
text = "I am so happy and excited!"
emotions = predict_emotions(text, model_path)
print(emotions)

Advanced Usage

import torch
from transformers import AutoTokenizer
import sys
sys.path.append(model_path)
from model import MultiLabelEmotionClassifier, load_model

# Load model manually
model, config = load_model(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path)

# Custom prediction with different threshold
def custom_predict(text, threshold=0.3):
    encoding = tokenizer(
        text,
        truncation=True,
        padding='max_length',
        max_length=128,
        return_tensors='pt'
    )
    
    model.eval()
    with torch.no_grad():
        logits = model(encoding['input_ids'], encoding['attention_mask'])
        probabilities = torch.sigmoid(logits)
        predictions = (probabilities > threshold).int()
    
    emotion_labels = ['amusement', 'anger', 'annoyance', 'caring', 'confusion', 'disappointment', 'disgust', 'embarrassment', 'excitement', 'fear', 'gratitude', 'joy', 'love', 'sadness']
    result = {emotion: {
        'predicted': bool(pred), 
        'probability': float(prob)
    } for emotion, pred, prob in zip(emotion_labels, predictions[0], probabilities[0])}
    return result

# Example with probabilities
result = custom_predict("I feel great today!", threshold=0.3)
print(result)

Model Architecture

Base: distilbert-base-uncased
Classification Head: Linear layer with dropout (dropout_rate=0.3)
Loss Function: BCEWithLogitsLoss
Activation: Sigmoid (for multilabel classification)

Training Details

Epochs: 3
Batch Size: 32
Learning Rate: 2e-05
Max Sequence Length: 128
Optimizer: AdamW with weight decay (0.01)
Scheduler: Linear warmup + decay

Files in this Repository

config.json: Model configuration
pytorch_model.bin: Model weights
tokenizer.json, tokenizer_config.json: Tokenizer files
model.py: Custom model class and utility functions
README.md: This file

Performance

Task: Multilabel Emotion Classification
Metrics: F1-Score (Micro & Macro), Accuracy
Validation Strategy: 80/20 train-validation split

Supported Emotions

amusement, anger, annoyance, caring, confusion, disappointment, disgust, embarrassment, excitement, fear, gratitude, joy, love, sadness

License

This model is released under the Apache 2.0 license.

Citation

@misc{firsttimeup2024,
  title={FirstTimeUp: Multilabel Emotion Classification Model},
  author={EnJiZ},
  year={2024},
  url={https://huggingface.co/EnJiZ/FirstTimeUp}
}

Contact

For questions or issues, please open an issue in the repository.

EnJiZ
/

FirstTimeUp