Instructions to use random1st/secguard-models with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use random1st/secguard-models with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="random1st/secguard-models",
	filename="secguard-guard.gguf",
)

llm.create_chat_completion(
	messages = "\"I like you. I love you\""
)

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use random1st/secguard-models with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf random1st/secguard-models
# Run inference directly in the terminal:
llama-cli -hf random1st/secguard-models

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf random1st/secguard-models
# Run inference directly in the terminal:
llama-cli -hf random1st/secguard-models

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf random1st/secguard-models
# Run inference directly in the terminal:
./llama-cli -hf random1st/secguard-models

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf random1st/secguard-models
# Run inference directly in the terminal:
./build/bin/llama-cli -hf random1st/secguard-models

Use Docker

docker model run hf.co/random1st/secguard-models

LM Studio
Jan
Ollama
How to use random1st/secguard-models with Ollama:
```
ollama run hf.co/random1st/secguard-models
```

Unsloth Studio new

How to use random1st/secguard-models with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for random1st/secguard-models to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for random1st/secguard-models to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for random1st/secguard-models to start chatting

Pi new

How to use random1st/secguard-models with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf random1st/secguard-models

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "random1st/secguard-models"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use random1st/secguard-models with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf random1st/secguard-models

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default random1st/secguard-models

Run Hermes

hermes

Docker Model Runner
How to use random1st/secguard-models with Docker Model Runner:
```
docker model run hf.co/random1st/secguard-models
```

Lemonade

How to use random1st/secguard-models with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull random1st/secguard-models

Run and chat with the model

lemonade run user.secguard-models-{{QUANT_TAG}}

List all available models

lemonade list

secguard-models / README.md

random1st

Fix MODEL_CARD placeholders (val loss, duration)

d5a36ff verified 19 days ago

preview code

raw

history blame contribute delete

2.99 kB

metadata

language: en
license: apache-2.0
tags:
  - security
  - shell-commands
  - safety-classifier
  - gguf
  - qwen3.5
datasets:
  - ML4Net/SSH-Shell-Attacks
  - TellinaTool/nl2bash
pipeline_tag: text-classification

secguard-guard — Shell Command Safety Classifier

Binary classifier for shell commands: safe vs destructive.

Model Details

Property	Value
Base model	Qwen/Qwen3.5-0.8B
Fine-tuning	LoRA (rank 16, α=32, 4.26M trainable / 752M total)
Quantization	Q8_0 (GGUF)
Size	~800 MB
Context	512 tokens
Inference	llama.cpp / llama-cpp-rs

Training

Dataset: 21,430 labeled examples (balanced 50/50, ChatML format)
- Destructive (10,715): SSH-Shell-Attacks honeypot commands (ML4Net, 408K sessions) + synthetic SaaS CLI patterns
- Safe (10,715): NL2Bash corpus (12.6K real admin commands) + synthetic dev/ops commands
Method: MLX LoRA, 16 layers, batch 4, lr 1e-5, 1000 iterations
Loss: Train 0.393, Val 0.401 (best at iter 400)
Test accuracy: 98.8% (500 held-out examples; precision 99.2%, recall 98.4%, F1 0.988)
Hardware: Apple Silicon M3 Max, ~30 minutes training

Notes on inference

Qwen3.5 reasoning models emit <think>…</think> blocks before the final answer. The runtime (secguard-brain) strips the thinking block via rfind("</think>") before matching the label, so the model is used as a classifier without retraining to suppress reasoning.

The MLX → GGUF pipeline requires three post-processing fixes for Qwen3.5 (tensor name rename, conv1d transpose, norm −1). Without them, the model produces multilingual token salad. This GGUF was produced through the fixed pipeline.

What it detects

Commands the model learns to classify as destructive:

File deletion (rm -rf, find -delete, shred)
Git history rewriting (push --force, reset --hard, rebase, filter-branch)
Database destruction (DROP TABLE, FLUSHALL, db.dropDatabase())
Cloud resource deletion (aws s3 rm, gcloud delete, terraform destroy)
Remote code execution (curl | bash, wget | sh)
Container/k8s cleanup (docker system prune, kubectl delete namespace)
SaaS destructive ops (stripe cancel, heroku apps:destroy)

Usage with secguard

This model is Phase 3 (ML brain) in secguard's three-phase guard:

Policy allowlist — known-safe commands (zero latency)
Heuristic rules — 40+ regex patterns (zero latency)
ML brain — this model (catches what rules miss)

secguard model     # downloads this GGUF to ~/.secguard/models/
secguard init --global     # installs Claude Code / Gemini / Codex hooks

Limitations

Trained on English commands only
SSH honeypot data doesn't represent all attack vectors
Confidence threshold: 0.85 (tunable in secguard config)
Below threshold → verdict falls through to safe (heuristic stays as backstop)

License

Apache 2.0