Emotion detection

A multi-label emotion detection project trained on GoEmotions with a DeBERTa v3 model. The repository includes the original training notebook and a Streamlit app for local inference.

What it does

• Predicts one or more emotions from text
• Uses a transformer classifier with sigmoid outputs
• Applies a tuned confidence threshold for multi-label prediction
• Shows detected emotions and confidence scores
• Displays a probability chart for all 28 GoEmotions labels

Tech stack

• Python
• Streamlit
• PyTorch
• Hugging Face Transformers
• pandas
• Altair

Project files

app.py                    Streamlit inference app
emotion_detection.ipynb   Training and evaluation notebook
label_map.json            GoEmotions label mapping
requirements.txt          Runtime dependencies
artifacts/                Local model files, ignored by Git

Run locally

Create and activate a virtual environment:

python -m venv .venv
.venv\Scripts\activate

Install dependencies:

pip install -r requirements.txt

Run the app:

streamlit run app.py

The app expects trained model files in artifacts/. Keep those files out of normal Git commits because the model checkpoint is large.

Training notebook

Open emotion_detection.ipynb in VS Code, Jupyter, or Colab to review the training workflow. The notebook covers dataset loading, preprocessing, transformer training, threshold tuning, final evaluation, per-label metrics, and misclassified examples.

Deployment

Use Streamlit Community Cloud or Hugging Face Spaces for a free recruiter demo.

For Hugging Face Spaces, create a Streamlit Space and upload:

• app.py
• requirements.txt
• label_map.json
• the model artifacts, preferably with Git LFS

Limitations

• The model checkpoint is too large for a normal Git commit
• Inference runs on CPU unless deployed with GPU hardware
• The app is a demo, not a moderation or clinical tool