This model card corresponds to CustomerGPT-0.5B-LoRA, a specialized customer support adapter fine-tuned on customer support conversations, intent classifications, and query enrichments. It is based on the lightweight Qwen/Qwen2.5-0.5B-Instruct base model, utilizing QLoRA (4-bit quantization + PEFT) for resource-friendly and fast execution.

GitHub Repository: CustomerGPT-FineTunning-1
Hugging Face Model Page: rudrakshrakeshzodage/CustomerGPT-0.5B-LoRA

📊 Dataset Specifications

The pipeline processes a raw dataset containing 26,872 support requests and expands it via data augmentation to enhance robustness.

1. Data Splits & Volume

Raw Records: 26,872
Cleaned & Filtered Records: 26,872 (100% pass rate at Quality Threshold)
Augmented Records: 107,488 (80,616 synthetic additions)

Split	Percentage	Records Count	Purpose
Train	80%	85,990	Model gradient updates
Validation	10%	10,748	Early stopping & hyperparameter tuning
Test	10%	10,750	Final evaluation

⚙️ Model Training Specifications

Parameter	Configuration / Specification
Base Model	`Qwen/Qwen2.5-0.5B-Instruct`
Training Method	QLoRA (4-bit Quantization + LoRA)
Quantization Scheme	NormalFloat4 (`nf4`), Double Quantization, `float16` compute type
LoRA Rank ($r$)	16
LoRA Alpha ($\alpha$)	32
LoRA Dropout	0.05
Target Modules	All linear layers (`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`)
Trainable Parameters	8,798,208 (1.75% of total 502,830,976)
Optimizer	`adamw_torch`
Learning Rate (LR)	2e-4 with Cosine Decay (ending at 0.0)
Effective Batch Size	8
Hardware	NVIDIA GeForce RTX 4060 Laptop GPU

📈 Fine-Tuning Metrics

Total Training Time: 2,293.04 seconds (~38.2 minutes)
Training Speed: 0.305 steps/second (~3.28 seconds/step)
Initial Loss: 2.4139
Final Training Loss: 0.6336
Average Training Loss: 0.8391

🚀 How to Load and Run Inference

You can load this LoRA adapter directly using the transformers and peft libraries:

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_name = "Qwen/Qwen2.5-0.5B-Instruct"
adapter_id = "rudrakshrakeshzodage/CustomerGPT-0.5B-LoRA"

# Load Tokenizer
tokenizer = AutoTokenizer.from_pretrained(base_model_name)

# Load Base Model
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Attach LoRA Adapter
model = PeftModel.from_pretrained(base_model, adapter_id)
model.eval()

# Chat Format Example
system_prompt = "You are CustomerGPT, a helpful and professional customer support assistant."
messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": "How to newsletter subscription"}
]

formatted_prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(formatted_prompt, return_tensors="pt").to("cuda")

with torch.no_grad():
    outputs = model.generate(
        **inputs, 
        max_new_tokens=256,
        temperature=0.7,
        top_p=0.9,
        repetition_penalty=1.1,
        pad_token_id=tokenizer.pad_token_id,
        eos_token_id=tokenizer.eos_token_id
    )
    
response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(response)