Emotion Detection Model for MindPadi (emotion_model)

This model is part of the MindPadi ecosystem — a mental health chatbot designed to offer empathetic, context-aware responses. emotion_model is a transformer-based sequence classification model trained to detect a range of emotional states from user input. It helps personalize chatbot responses by understanding the emotional tone of each message.

🧠 Model Summary

• Task: Emotion Classification
• Architecture: Transformer-based (likely BERT or DistilBERT)
• Labels: happy, sad, angry, neutral, fearful, disgust, surprised, etc.
• Framework: 🤗 Transformers (PyTorch backend)
• Use Case: Core emotion recognition module in app/chatbot/emotion.py

🧾 Intended Use

✔️ Primary Use Cases

• Detect user emotions in chat messages.
• Adjust response tone and therapy prompts in MindPadi.
• Support emotional trend tracking in mood analytics.

🚫 Not Recommended For

• Clinical diagnosis or treatment decisions.
• Emotion detection in highly formal or technical language (e.g., legal, medical).
• Non-English inputs (English-only training data).

📚 Training Details

• Training Script: training/train_emotion_model.py
• Datasets: A mix of publicly available emotion corpora (e.g., GoEmotions) and proprietary datasets stored in training/datasets/
• Preprocessing:
  • Cleaned for offensive language and class imbalance.
  • Tokenized using AutoTokenizer from Hugging Face Transformers.
• Hyperparameters:
  • Epochs: ~4–6
  • Batch Size: 16–32
  • Learning Rate: 2e-5 to 3e-5
• Loss Function: CrossEntropyLoss
• Optimizer: AdamW

✅ Evaluation

• Metrics: Accuracy, F1-score (micro, macro), confusion matrix
• Evaluation Script: training/evaluate_model.py
• Performance:
  • Accuracy: ~87%
  • Macro F1: ~85%
  • Robust across common emotional states like sad, happy, angry
• Visualization: See lstm_accuracy_bert.png for comparisons

📦 Files

The model directory includes:

File	Purpose
config.json	Model architecture configuration
model.safetensors	Trained model weights
tokenizer.json, vocab.txt	Tokenizer config
merges.txt (if BPE-based)	Byte-pair encoding rules
checkpoint-*/ (optional)	Intermediate training checkpoints

🔍 Example Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "mindpadi/emotion_model"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

text = "I feel so overwhelmed and tired."
inputs = tokenizer(text, return_tensors="pt")
outputs = model(**inputs)
predicted_class = torch.argmax(outputs.logits, dim=1).item()

print("Predicted emotion class:", predicted_class)

💡 Integration

Integrated in:

• app/chatbot/emotion.py: Emotion detection during each chat turn.
• app/utils/analytics.py: Aggregates emotions for weekly mood charts.
• LangGraph: Used in flow state personalization nodes.

⚠️ Limitations

• Bias: May inherit cultural or gender biases from training data.
• Language: English only.
• False Positives: Sarcasm or ambiguous text may confuse predictions.
• Not Clinical: Should not be relied upon for medical-level emotional assessments.

🧑‍⚖️ Ethical Considerations

• MindPadi informs users that they are interacting with AI.
• Emotion analysis is used only to guide and personalize chatbot responses.
• All usage must respect user privacy (see app/tools/encryption.py for encryption methods).

🧩 License

MIT License. You are free to use, modify, and distribute the model with attribution.

📬 Contact

• Project: MindPadi Mental Health Chatbot
• Author: MindPadi Team
• Email: [[email protected]]
• GitHub: [github.com/mindpadi]

Last updated: May 2025