hinglish-intent-classifier

🎙️ Hinglish Intent Classifier (LoRA + DistilBERT)

Fine-tuned DistilBERT Multilingual adapter using LoRA (PEFT) for Intent Classification on Code-Mixed Hindi-English (Hinglish) conversational utterances. Built for voice-agent Natural Language Understanding (NLU) pipelines.

🎯 Target Intent Classes

1. complaint (e.g., service issues, refund delays)
2. purchase_inquiry (e.g., catalog questions, warranty inquiries)
3. price_negotiation (e.g., asking for discounts, coupon requests)
4. callback_request (e.g., busy right now, call later)
5. not_interested (e.g., refusal, DND requests)
6. positive_confirmation (e.g., deal acceptance, confirmation)

📊 Benchmark Results

🚀 Quick Start & Usage

import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from peft import PeftModel

base_model_name = "distilbert-base-multilingual-cased"
adapter_repo = "yashasvijadav03/hinglish-intent-classifier"

# Intent labels mapping
id2label = {
    0: "complaint",
    1: "purchase_inquiry",
    2: "price_negotiation",
    3: "callback_request",
    4: "not_interested",
    5: "positive_confirmation"
}

# 1. Load Tokenizer & Base Model
tokenizer = AutoTokenizer.from_pretrained(base_model_name)
base_model = AutoModelForSequenceClassification.from_pretrained(
    base_model_name,
    num_labels=len(id2label),
    id2label=id2label,
)

# 2. Attach Fine-Tuned LoRA Adapter
model = PeftModel.from_pretrained(base_model, adapter_repo)
model.eval()

# 3. Classify an Utterance
text = "Thoda discount de do na, price bohot zyada hai."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)

with torch.no_grad():
    logits = model(**inputs).logits
    probs = F.softmax(logits, dim=-1)

predicted_id = torch.argmax(probs, dim=-1).item()
print(f"Predicted Intent: {id2label[predicted_id]} (Confidence: {probs[0][predicted_id]:.2%})")
# Output: Predicted Intent: price_negotiation (Confidence: 98.40%)

🛠️ Training Specifications

• Base Model: distilbert-base-multilingual-cased (135M params)
• Trainable Parameters: 1,185,798 (~0.87% of total model params via LoRA)
• LoRA Parameters: Rank $r=16$, Alpha $\alpha=32$, Dropout $0.1$, Target Modules: q_lin, v_lin
• Training Epochs: 3
• Learning Rate: 5e-4 with Linear Warmup