learn-abc/magicSupport-intent-classifier

MagicSupport Intent Classifier (BERT Fine-Tuned)

Overview

This model is a fine-tuned bert-base-uncased model for multi-class intent classification in customer support environments.

It is optimized for:

• Fast inference
• High accuracy
• Low deployment cost
• Production-ready intent routing for support systems

The model is designed for the MagicSupport platform but is generalizable to structured customer support intent detection tasks.

---

Model Details

• Base Model: bert-base-uncased
• Architecture: BertForSequenceClassification
• Task: Multi-class intent classification
• Number of Intents: 28
• Training Dataset: bitext/Bitext-customer-support-llm-chatbot-training-dataset
• Loss: CrossEntropy with class weights
• Framework: Hugging Face Transformers (PyTorch)

---

Performance

Validation Metrics (Epoch 5)

• Accuracy: 0.9983
• F1 Micro: 0.9983
• F1 Macro: 0.9983
• Validation Loss: 0.0087

The model demonstrates strong generalization and stable convergence across 5 epochs.

---

Example Predictions

Query	Predicted Intent	Confidence
I want to cancel my order	cancel_order	0.999
How do I track my shipment	delivery_options	0.997
I need a refund for my purchase	get_refund	0.999
I forgot my password	recover_password	0.999
I have a complaint about your service	complaint	0.996
hello	FALLBACK	0.999

The model also correctly identifies low-information inputs and maps them to a fallback intent.

---

Intended Use

This model is intended for:

• Customer support intent classification
• Chatbot routing
• Support ticket categorization
• Voice-to-intent pipelines (after STT)
• Pre-routing before LLM or RAG systems

Typical production flow:

User Query → BERT Intent Classifier → Route to:

• Knowledge Base Retrieval
• Ticketing System
• Escalation to Human
• Fallback LLM

---

Example Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model and tokenizer from HuggingFace Hub
model_name = "learn-abc/magicSupport-intent-classifier"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Set device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
model.eval()

# Prediction function
def predict_intent(text, confidence_threshold=0.75):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=64)
    inputs = {k: v.to(device) for k, v in inputs.items()}
    
    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs.logits
        probs = torch.softmax(logits, dim=-1)
        confidence, prediction = torch.max(probs, dim=-1)
    
    predicted_intent = model.config.id2label[prediction.item()]
    confidence_score = confidence.item()
    
    # Apply confidence threshold
    if confidence_score < confidence_threshold:
        predicted_intent = "FALLBACK"
    
    return {
        "intent": predicted_intent,
        "confidence": confidence_score
    }

# Example usage
queries = [
    "I want to cancel my order",
    "How do I track my package",
    "I need a refund",
    "hello there"
]

for query in queries:
    result = predict_intent(query)
    print(f"Query: {query}")
    print(f"Intent: {result['intent']}")
    print(f"Confidence: {result['confidence']:.3f}\n")

---

Design Decisions

• BERT selected over larger LLMs for:

  • Low latency
  • Cost efficiency
  • Predictable inference
  • Edge deployability

• Class weighting applied to mitigate dataset imbalance.

• High confidence outputs indicate strong separation between intent classes.

---

Known Limitations

• Designed for structured customer support queries.

• May struggle with:

  • Highly conversational multi-turn context
  • Extremely domain-specific enterprise terminology
  • Heavy slang or multilingual input

• Not trained for open-domain conversation.

---

Future Improvements

• Add MagicSupport real production data for domain adaptation.
• Add hierarchical intent structure.
• Introduce confidence threshold calibration.
• Add OOD (Out-of-Distribution) detection.
• Quantized inference version for edge deployment.

---

License

Specify your intended license here (e.g., MIT, Apache-2.0).

---

Citation

If using this model in research or production, please cite appropriately.

---

Model Card Author

For any inquiries or support, please reach out to:

• Author: Abhishek Singh
• LinkedIn: My LinkedIn Profile
• Portfolio: Abhishek Singh Portfolio