llmshield-1b-finetuned

---

🧠 llmshield-1b-finetuned — Fine-Tuned Llama 3.2 Instruction Model

✅ Model Overview

LLMShield is a lightweight, efficient, instruction-following language model based on:

• Base Model: unsloth/Llama-3.2-1B-Instruct
• Fine-tuning Method: Supervised Fine-Tuning (SFT)
• Parameter Efficiency: LoRA adapters using PEFT
• Training Toolkit: TRL SFTTrainer + Unsloth chat utilities
• Intended Use: Answer questions concisely, follow instructions, general purpose conversation

The model is optimized for:

• Direct answers
• Instruction following
• Short explanations
• Simple reasoning
• Code explanation (small scale)

Because it is only 1B parameters, it runs extremely fast on CPU and mobile hardware.

---

🛠️ Training Details

Fine-tuning Strategy

The model was trained using:

• SFTTrainer (supervised fine tuning)
• LoRA adapters (very low VRAM cost)
• Training on assistant responses only
• Using Llama-style chat formatting
• No packing (stable training)

---

🚀 How To Use This Model

Below is the official working inference code, verified to run on Colab / local Jupyter / any Python environment.

---

✅ Install Requirements

pip install transformers accelerate peft torch huggingface_hub

---

✅ Load Model & Tokenizer

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

# notebook_login()   # Required only if model is private

# Base Llama 3.2 model
base_model = "unsloth/Llama-3.2-1B-Instruct"

# Your fine tuned LoRA adapter
ft_model = "arif-butt/llmshield-1b-finetuned"

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(base_model)

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Apply LoRA adapter
model = PeftModel.from_pretrained(base_model, ft_model)

# Evaluation mode
model.eval()

---

✅ Chat Function Example

def chat(user_message, max_new_tokens=80):

    # Construct structured messages required by Llama chat template
    messages = [
        {
            "role": "system",
            "content": (
                "Answer the user's question directly and concisely. "
                "Do NOT ask new questions or create a Q&A list."
            )
        },
        {"role": "user", "content": user_message},
    ]

    # Convert messages → tokens using built-in template
    inputs = tokenizer.apply_chat_template(
        messages,
        add_generation_prompt=True,
        return_tensors="pt",
    ).to(model.device)

    # Generate continuation
    outputs = model.generate(
        inputs,
        max_new_tokens=max_new_tokens,
        temperature=0.7,
        top_p=0.9,
        do_sample=True,
    )

    # Remove prompt → decode assistant reply
    generated_tokens = outputs[0, inputs.shape[-1]:]
    answer = tokenizer.decode(generated_tokens,
                              skip_special_tokens=True).strip()
    
    return answer

---

✅ Example Usage

print(chat("What is LLMShield?"))

Example Output:

LLMShield is a fine-tuned Llama 3.2 1B instruction model designed
to answer questions concisely and follow user instructions.

---

⭐ Capabilities

This model performs well at:

• Question answering
• Instruction following
• Short summaries
• Explaining concepts simply
• General conversation
• Running on low hardware (CPU / mobile)

---

⚠️ Limitations

As a small LLM, it may:

• Produce incorrect facts
• Have weak long reasoning
• Lack knowledge beyond training date
• Produce unsafe or biased responses (always verify)

Do not use it for:

• Medical advice
• Legal advice
• Financial advice
• Anything requiring guaranteed correctness

---

🧪 Intended Users

• Students learning LLMs
• Developers building lightweight apps
• Educators teaching fine-tuning
• Anyone wanting fast local inference

---

📦 Repository Files

The model card applies to a repo containing at least:

• adapter_model.safetensors or pytorch_lora_weights.bin
• tokenizer.json
• config.json
• special_tokens_map.json
• README.md (this file)

---

📄 License

Specify your license:

Example:

MIT License

or:

Apache-2.0

---

📎 Citation

If you use this model in research:

LLMShield: Fine-tuned Llama 3.2 1B model
Author: Dr. Muhammad Arif Butt, Um E Abeeha, Khalood Sami, and Alisha Shahid.
HuggingFace: arif-butt/llmshield-1b-finetuned