| --- |
| library_name: transformers |
| tags: |
| - peft |
| - qlora |
| - lora |
| - cybersecurity |
| - text-generation |
| base_model: microsoft/Phi-3-mini-4k-instruct |
| --- |
| |
| # Model Card for Cybersecurity Assistant Sandbox |
|
|
| This model is a proof-of-concept conversational LLM fine-tuned to provide clean, direct definitions and concept explanations for machine learning architectures, NLP principles, and cybersecurity fundamentals. |
|
|
| ## Model Details |
|
|
| ### Model Description |
|
|
| This is a Parameter-Efficient Fine-Tuning (PEFT) adapter layer built using Low-Rank Adaptation (LoRA). It has been adapted from a quantized 4-bit base LLM to deliver specific, highly structured answers to technical instructions while minimizing verbose internet filler text. |
|
|
| - **Developed by:** bixby404 |
| - **Model type:** Causal Language Model (LoRA Adapter) |
| - **Language(s) (NLP):** English |
| - **Finetuned from model:** microsoft/Phi-3-mini-4k-instruct |
|
|
| ### Model Sources |
|
|
| - **Repository:** https://huggingface.co/bixby404/cybersecurity-assistant |
| - **Demo:** Hosted via Google Colab and Gradio UI Sandbox |
|
|
| ## Uses |
|
|
| ### Direct Use |
|
|
| This model is intended to be used directly as a lightweight technical assistant. It excels at answering explicit, structured definitions matching the instructions provided in its training regimen. |
|
|
| ### Out-of-Scope Use |
|
|
| This model is **not** designed for: |
| - Writing production-grade exploit scripts or malicious software. |
| - Analyzing enterprise system log traffic in real-time. |
| - Serving as a standalone automated incident response tool without human oversight. |
|
|
| ## Bias, Risks, and Limitations |
|
|
| Due to the compact dataset size (100 rows), the model acts primarily as a behavior/style filter rather than an extensive new knowledge repository. It carries a structural dependency on its base architecture (`Phi-3`) for general reasoning and might display style regression if prompted with heavily out-of-domain scenarios. |
|
|
| ### Recommendations |
|
|
| Users should verify any specialized security remediations or definitions generated against official standards (such as NIST or MITRE ATT&CK frameworks) before executing them in live corporate lab environments. |
|
|
| ## How to Get Started with the Model |
|
|
| Use the PyTorch execution code below in a standard Google Colab environment containing a T4 GPU runtime to interface with this model adapter: |
|
|
| ```python |
| import torch |
| from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig, pipeline |
| from peft import PeftModel |
| |
| base_model_name = "microsoft/Phi-3-mini-4k-instruct" |
| hf_adapter_id = "bixby404/cybersecurity-assistant" |
| |
| bnb_config = BitsAndBytesConfig( |
| load_in_4bit=True, |
| bnb_4bit_quant_type="nf4", |
| bnb_4bit_compute_dtype=torch.float16 |
| ) |
| |
| model = AutoModelForCausalLM.from_pretrained( |
| base_model_name, |
| quantization_config=bnb_config, |
| device_map="auto" |
| ) |
| |
| model = PeftModel.from_pretrained(model, hf_adapter_id) |
| tokenizer = AutoTokenizer.from_pretrained(base_model_name) |
| |
| generator = pipeline("text-generation", model=model, tokenizer=tokenizer) |
| test_prompt = "### Instruction:\nWhat is the primary function of an LLM?\n\n### Response:\n" |
| |
| outputs = generator(test_prompt, max_new_tokens=50, return_full_text=False) |
| print(outputs[0]['generated_text']) |
