---
language:
- hi
- ta
- en
tags:
- customer-support
- hinglish
- tanglish
- fine-tuned
- qwen2.5
- qlora
base_model: Qwen/Qwen2.5-3B-Instruct
pipeline_tag: text-generation
---

# SupportLM — Hinglish/Tanglish Customer Support Model

A fine-tuned Qwen2.5-3B model for Indian e-commerce customer support
in code-switched language (Hinglish and Tanglish).

## What it does

Given a customer support message in Hinglish or Tanglish, the model:
1. Classifies the issue type into one of 8 categories
2. Extracts the order ID if present in the message
3. Generates an empathetic response in the same language the customer used

## Benchmark Results

Evaluated on 159 held-out test examples never seen during training.

| Metric | GPT-4o-mini | SFT Model | DPO Model |
|--------|-------------|-----------|-----------|
| Classification accuracy | 74.8% | **79.9%** | 78.0% |
| Order ID extraction | 98.7% | **99.4%** | 99.4% |
| Language match rate | 96.2% | 99.4% | **100.0%** |

SFT model chosen for deployment — best classification accuracy.

## Issue Types

order_status, return_request, payment_issue, delivery_complaint,
product_defect, cancellation_request, wrong_item, discount_query

## Training Details

- Base model: Qwen/Qwen2.5-3B-Instruct
- Training data: 1277 synthetic Hinglish/Tanglish examples
- Training method: QLoRA (r=16, alpha=32) + DPO (beta=0.1)
- Hardware: Google Colab T4 GPU (free tier)
- Training time: ~40 minutes total (SFT + DPO)

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import json

model = AutoModelForCausalLM.from_pretrained("GuruVarshini/supportlm-qwen2.5-3b")
tokenizer = AutoTokenizer.from_pretrained("GuruVarshini/supportlm-qwen2.5-3b")

message = "bhai mere order ka kya hua yaar 3 din ho gaye"

messages = [
    {"role": "system", "content": "You are a customer support agent for an Indian e-commerce platform. Given a customer message, respond with a JSON object containing issue_type, order_id, and response."},
    {"role": "user", "content": message}
]

input_ids = tokenizer.apply_chat_template(
    messages,
    return_tensors="pt",
    add_generation_prompt=True
)
output = model.generate(input_ids, max_new_tokens=200)
print(tokenizer.decode(output[0][input_ids.shape[1]:], skip_special_tokens=True))