Modelpublic

erkanrzgcc/octopus-v0.8.1

sourceHugging Facegemmaupdated 2mo agoView on Hugging Face

0likes22downloads

Model Card

Octópus — v0.8.1

An agentic LLM for cybersecurity. Octópus emits structured tool calls that an agent runtime executes against real security tools, behind an authorization gate. It covers red team (recon, exploitation) and blue team (detection, hardening, incident response) across network and Linux, and introduces itself as "Ben Octópus".

🐙 Project & agent harness: https://github.com/erkanrzgc/octopus

This is the current default release (persona v0.3). It ships the bf16 LoRA adapter plus a ready-to-run GGUF (Q4_K_M) quantization.

Run locally (Ollama)

bash

huggingface-cli download erkanrzgcc/octopus-v0.8.1 octopus-v81-Q4_K_M.gguf --local-dir .
# Modelfile (RAW): FROM ./octopus-v81-Q4_K_M.gguf  +  PARAMETER stop "<end_of_turn>"
ollama create octopus-v81 -f Modelfile
ollama run octopus-v81 "Merhaba, kendini tanıt."

Files

File	Description
`adapter_model.safetensors`	bf16 LoRA adapter (r=32, α=32)
`octopus-v81-Q4_K_M.gguf`	merged + quantized weights for local inference

Training

—Method: bf16 LoRA (r=32, α=32, 7 target modules) — no 4-bit (it corrupts the merged base).
—Engine: transformers + peft + TRL.
—Data: security Q&A + 117-tool agentic use + reasoning/persona seeds.

⚠️ Responsible Use

Authorized, legal, lab / CTF / educational use only. Octópus assists only on systems the operator is explicitly authorized to test and refuses attack steps against unauthorized real targets. It must not be used for unauthorized access or harm; responsibility rests entirely with the user.

License

Weights are governed by the base model's Terms of Use. Project code is MIT (see the GitHub repository).