Shlok307/gemma-3-1b-it-cyber-lora

Model Card for Model ID

This model is a LoRA adapter fine-tuned on top of google/gemma-3-1b-it, optimized for safe cybersecurity explanations, vulnerability analysis, and general security guidance. It improves the base Gemma model’s performance on technical cyber topics while keeping responses safe, helpful, and beginner-friendly.

Model Description

This is a lightweight LoRA (QLoRA) adapter trained on a curated dataset of cybersecurity conversations containing system, user, and assistant messages.
The fine-tuning process enhances Gemma-3-1B to be:

• Better at explaining security concepts
• More fluent in cybersecurity terminology
• More structured and coherent in its answers
• More helpful, safer, and suitable for educational use

The underlying base model remains google/gemma-3-1b-it.

Credits

• Developed by: Shlok Talhar
• Finetuned from: google/gemma-3-1b-it
• Model type: LoRA adapter for causal language modeling
• Language: English
• License: Gemma License
• Training method: QLoRA using PEFT

Model Sources

• Repository: https://huggingface.co/Shlok307/gemma-3-1b-it-cyber-lora
• Base Model: https://huggingface.co/google/gemma-3-1b-it

How to Get Started with the Model

Use the code below to get started with the model.

!pip install -U bitsandbytes accelerate

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

base = "google/gemma-3-1b-it"
adapter = "Shlok307/gemma-3-1b-it-cyber-lora"

# 4-bit quantization (recommended)
quant_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype="float16",
    bnb_4bit_quant_type="nf4"
)

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(base)

# Load base model in 4-bit
model = AutoModelForCausalLM.from_pretrained(
    base,
    quantization_config=quant_config,
    device_map="auto"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(model, adapter)
model.eval()

# Example prompt
prompt = (
    "System: You are a straight forward cybersecurity and Fraud detection assistant.\n"
    "User: Explain how SQL injection works on website.\n"
    "Assistant:"
)

# Tokenize
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

# Inference
with torch.no_grad():
    output = model.generate(
        **inputs,
        max_new_tokens=450,
        temperature=0.7
    )

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

📊 Model Parameters

• Total parameters: 656,968,832
• LoRA trainable parameters: 5,963,776
• Percentage of model fine-tuned: 0.9078%
• Parameter-efficient tuning: Yes (LoRA / QLoRA)

Prompt Format

System: <system>
User: <user>
Assistant: <assistant>

Summary of Improvements

• Strong improvements in concept explanations
• Better consistency and clarity on cybersecurity tasks
• Safe outputs with reduced harmful behavior