Maintainer: Dr. Sabri Sallani Expertise: AI Research & Cybersecurity Adapter type: LoRA (Low-Rank Adaptation) Base model: mistralai/Mistral-7B-v0.1 (FP16) Intended use: Offline edge deployment for CyberAI & Blue-Team scenarios License: Apache 2.0 (see LICENSE)

📖 Overview

ELISARCyberAIEdge7B is a LoRA adapter crafted by Dr. Sabri Sallani—AI & cybersecurity researcher—to specialize Mistral-7B for offline, on-device CyberAI and “Blue AI” (defensive) applications. Once merged with the FP16 base, you obtain a single ~5 GB GGUF that runs natively on edge hardware (e.g., Raspberry Pi 4, Jetson Nano, NVIDIA T4) without internet access.

Key points:

🔧 LoRA-only: Contains low-rank delta-weights for Mistral-7B.
🛠️ Edge-optimized: Full merged GGUF runs entirely offline on typical edge GPUs/accelerators.
🚀 Cybersecurity focus: Fine-tuned on “ELISAR CyberAI Edge” corpus—vulnerability descriptions, incident reports, secure-coding examples, threat intelligence summaries.
👤 Authored by Dr. Sabri Sallani: Published under the ELISAR initiative.

⚙️ Installation

Python dependencies

pip install transformers peft accelerate sentencepiece torch
(Optional) llama.cpp + GGUF tools (to merge and run offline)

bash
1# Clone and install gguf-py 2git clone --depth 1 https://github.com/ggml-org/llama.cpp.git 3pip install ./llama.cpp/gguf-py 4pip install llama-cpp-python

→ Use these tools to merge LoRA + base weights into a single GGUF.

🐍 Usage

1. Inference with `transformers` + `PEFT` (online GPU/CPU)

python

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

BASE_ID    = "mistralai/Mistral-7B-v0.1"
ADAPTER_ID = "sallani/ELISARCyberAIEdge7B"

# 1) Load Mistral-7B base (FP16 or BF16) with automatic device placement
tokenizer = AutoTokenizer.from_pretrained(BASE_ID, use_fast=True)
base_model = AutoModelForCausalLM.from_pretrained(
    BASE_ID,
    torch_dtype="auto",
    device_map="auto"
)

# 2) Load LoRA adapter on top
model = PeftModel.from_pretrained(
    base_model,
    ADAPTER_ID,
    torch_dtype="auto",
    device_map="auto"
)
model.eval()

# 3) Perform inference
prompt = (
    "### Instruction:\n"
    "Propose a set of secure-coding guidelines for Python web applications.\n"
    "### Response:\n"
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    out = model.generate(
        **inputs,
        max_new_tokens=128,
        temperature=0.8,
        top_p=0.9,
        repetition_penalty=1.1
    )

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

device_map="auto" places weights on GPU/CPU automatically (FP16 when supported).
Adjust sampling parameters (temperature, top_p, repetition_penalty) for your use case.

2. Offline Edge Deployment via `llama.cpp` (Merged GGUF)

Merge LoRA + base into a single GGUF

bash
1python3 llama.cpp/convert_lora_to_gguf.py \ 2 sallani/ELISARCyberAIEdge7B \ # LoRA repo or local folder 3 --base-model-id mistralai/Mistral-7B-v0.1 \ # HF ID of FP16 base 4 --outfile elisar_full_f16.gguf # Output GGUF (~5 GB)
- The script pulls the FP16 base automatically from HF, applies LoRA deltas, and writes a merged GGUF.
Run inference on edge
- Copy elisar_full_f16.gguf to your edge device (Jetson Nano, Raspberry Pi 4 + GPU, NVIDIA T4).
- Use llama.cpp binary to run:
  
  bash
  1./llama.cpp/main \ 2 -m elisar_full_f16.gguf \ 3 -p "### Instruction: Audit the following log entries for suspicious activity.\n---\n<log lines>\n---\n### Response:" \ 4 --temp 0.7 \ 5 --repeat_penalty 1.1 \ 6 --n 128
- No internet is required once the GGUF is on-device.

📐 Model Details

Base architecture: Mistral-7B-v0.1 (40 transformer layers, 4096-dim embedding, 32 heads, causal LM).
LoRA configuration:
- Rank = 64, α = 16
- Applied to Q/K/V and feed-forward projections
- Adapter snapshots ≈ 168 MB
Training corpus (ELISAR CyberAI Edge):
- Public vulnerability databases (CVE entries, CVSS scoring).
- Real-world incident reports (MITRE ATT&CK red vs. blue logs).
- Secure-coding patterns (OWASP Top 10, SAST examples).
- Blue-team playbooks and defensive strategies.
Hyperparameters:
- Learning rate = 1e-4, batch size = 16 per GPU, 3 epochs on 8×A100 (FP16).
- Validation on unseen CVE descriptions and red-team prompts.
Merged GGUF (FP16):
- ~5 GB total after merging and trimming unnecessary metadata for on-device use.

🔖 Prompt Guidelines

Structured prompt

### Instruction:
<clear cybersecurity or defensive AI task>
### Response:

Recommended sampling
- temperature=0.7–0.9 for balanced creativity.
- top_p=0.9 for nucleus sampling.
- repetition_penalty=1.1 to reduce loops.

language: en license: apache-2.0 tags:

gguf
quantized
cybersecurity
edge-llm
lora
mistral
elisar model_name: ELISARCyberAIEdge7B-LoRA-GGUF pipeline_tag: text-generation datasets:
custom widget:
text: "What are the main threats targeting OT environments?"

📊 Stats & Adoption

🔄 Download tracking: Enabled
📥 Total downloads (last 30 days): auto-updated by HF
🧪 Being tested on:
- Jetson Nano (Ubuntu 20.04, CUDA 11.4)
- Raspberry Pi 4 (with Coral TPU)
- NVIDIA T4 + LLaMA.cpp

Want to share your benchmarks? Open an Issue or pull request.

⚠️ License & Citation

License: Apache 2.0 (see LICENSE).
Attribution:

Sallani, S. (2025). ELISARCyberAIEdge7B: LoRA adapter for Mistral-7B specializing in offline CyberAI Edge tasks. Hugging Face Model Hub: sallani/ELISARCyberAIEdge7B.

🛠️ Support & Contact

Report issues or feature requests: https://huggingface.co/sallani/ELISARCyberAIEdge7B/issues
Contact the author: Dr. Sabri Sallani • GitHub: @sallani • Email: sabri.sallani@cyberaiedge.com • LinkedIn: linkedin.com/in/sabri-sallani

Thank you for using ELISARCyberAIEdge7B. This adapter empowers secure, offline AI at the edge for next-gen CyberAI and Blue-Team applications.

sallani/ELISARCyberAIEdge7B

ELISARCyberAIEdge7B