Qwen3-8B-LoRA-Merged 🧠

This is a QLoRA fine-tuned version of the Qwen/Qwen3-8B model.
It was fine-tuned using 4-bit quantization (NF4) and LoRA adapters for efficient domain adaptation.

---

🧩 Model Details

• Base Model: Qwen/Qwen3-8B
• Fine-tuning Technique: QLoRA (Low-Rank Adaptation)
• Quantization: 4-bit NF4, bfloat16 compute
• Adapter Rank (r): 64
• LoRA α (alpha): 128
• Dropout: 0.1
• Context Length: 1024 tokens
• Training Steps: 150
• Optimizer: AdamW
• Learning Rate: 5e-5

---

💬 System Prompt

You are a concise and expert AI assistant specializing in fine-tuning, quantization, and efficient model training.
Always explain concepts clearly, use technical precision, and provide short code examples when useful.

---

🧠 Example Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "varunpruthviraj/qwen3-8b-lora-merged"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

prompt = "Explain how QLoRA differs from traditional fine-tuning."
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=200)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))