DistilBERT for Intent Classification

Overview

• Architecture: DistilBERT (distilbert-base-uncased) for sequence classification
• Task: Single-label intent classification of HR queries using merged user query and context
• Dataset: ~133 samples, 12 intent classes, 80/20 train/validation split

Training Details

• Epochs: 5
• Batch Size: 8
• Learning Rate: 5e-5
• Optimizer: AdamW
• Loss: CrossEntropyLoss

Evaluation Metrics (Validation Set)

Metric	Value
Accuracy	88.89%
Precision	100%
Recall	88.89%
Loss	1.4586

Usage Example

text = "Share offer with Santhosh [context: {domain: HR, topic: onboarding, subject: offer letter}]" inputs = tokenizer(text, return_tensors='pt', truncation=True, padding=True) with torch.no_grad(): logits = model(**inputs).logits pred_id = logits.argmax(dim=1).item()

Comments

• Consistent strong results on validation set.
• Model is robust for HR chatbot/automation intent tasks.
• Consider more data or further tuning for additional improvement.

---

For best results, ensure your production inference pipeline preprocesses and tokenizes input exactly as done for the training data.

---

In summary:
You’ve followed the right steps for distilbert-based intent classification and your documentation—combined with this detailed evaluation/usage section—will be clear and informative for anyone using your model!