mindpadi/intent_fallback

Intent Fallback Classifier (MindPadi)

This model serves as a lightweight fallback intent classifier for the MindPadi mental health assistant. It is designed to handle ambiguous or unrecognized user inputs when the primary intent classifier fails or yields low confidence. It helps maintain robustness in routing user messages to the appropriate modules in the chatbot backend.

🧠 Model Summary

• Model Type: Transformer-based classifier (BERT or DistilBERT variant)
• Task: Text classification (intent prediction)
• Primary Purpose: Backup routing in case of low-confidence from main classifier
• Size: Lightweight (optimized for low-latency inference)
• Files:
  • config.json
  • pytorch_model.bin or model.safetensors
  • tokenizer.json, vocab.txt

🧾 Intended Use

✔️ Use Cases

• Predicting fallback intents like "unknown", "help", "clarify", etc.
• Activating clarification routines or default flows when the main intent classifier is uncertain
• Used in app/chatbot/intent_classifier.py as fallback_model.predict(...)

🚫 Not Recommended For

• Serving as the primary intent classifier for all inputs
• Handling highly nuanced or multi-intent queries
• Clinical or domain-specific intent disambiguation

🏋️‍♀️ Training Details

• Dataset: Internal fallback samples and "unknown intent" examples

  • Location: training/datasets/fallback_intents.json
  • Includes mislabeled, ambiguous, or noisy utterances

• Script: training/train_intent_classifier.py (with fallback mode enabled)

• Classes:

  • "unknown"
  • "clarify"
  • "greeting"
  • "exit"
  • "irrelevant"
  • "default"

Hyperparameters

• Model: bert-base-uncased or similar
• Batch Size: 16
• Learning Rate: 3e-5
• Epochs: 3–4
• Max Sequence Length: 128

📈 Evaluation

• Accuracy: ~91% on test set of ambiguous intent queries
• F1-Score (unknown intent): 0.94
• Confusion Matrix: Available in logs/fallback_intent_eval.png
• Performance Benchmark: Inference latency < 50ms on CPU

💬 Example Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

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

inputs = tokenizer("I'm not sure what I need", return_tensors="pt")
outputs = model(**inputs)
predicted_class = torch.argmax(outputs.logits, dim=1)

print(predicted_class.item())  # e.g., 0 => "unknown"

🔧 Integration in MindPadi

This model is used as:

• A safety net in app/chatbot/intent_classifier.py
• Part of the router fallback in app/chatbot/intent_router.py
• Optional validation in LangGraph workflows for ambiguous queries

⚠️ Limitations

• May overgeneralize rare intents as "unknown"
• Trained on a relatively small fallback dataset
• May require manual thresholds for activation in hybrid systems
• English-only

🧪 Deployment (Optional)

For real-time inference via Hugging Face Inference Endpoints:

import requests

api_url = "https://<your-endpoint>.hf.space"
headers = {"Authorization": f"Bearer <your-token>", "Content-Type": "application/json"}
payload = {"inputs": "I'm not sure what I need"}

response = requests.post(api_url, headers=headers, json=payload)
print(response.json())

📜 License

MIT License – Open for personal and commercial use with attribution.

📬 Contact

• Project: MindPadi AI
• Team: MindPadi Developers
• Email: [[email protected]]
• GitHub: [https://github.com/mindpadi]

Last updated: May 2025