Instructions to use virgiliolrf2/bagley-v10 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use virgiliolrf2/bagley-v10 with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("/vol/foundation-sec-8b")
model = PeftModel.from_pretrained(base_model, "virgiliolrf2/bagley-v10")

llama-cpp-python

How to use virgiliolrf2/bagley-v10 with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="virgiliolrf2/bagley-v10",
	filename="bagley-v10-Q4_K_M.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use virgiliolrf2/bagley-v10 with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf virgiliolrf2/bagley-v10:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf virgiliolrf2/bagley-v10:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf virgiliolrf2/bagley-v10:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf virgiliolrf2/bagley-v10:Q4_K_M

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 virgiliolrf2/bagley-v10:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf virgiliolrf2/bagley-v10:Q4_K_M

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 virgiliolrf2/bagley-v10:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf virgiliolrf2/bagley-v10:Q4_K_M

Use Docker

docker model run hf.co/virgiliolrf2/bagley-v10:Q4_K_M

LM Studio
Jan

vLLM

How to use virgiliolrf2/bagley-v10 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "virgiliolrf2/bagley-v10"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "virgiliolrf2/bagley-v10",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/virgiliolrf2/bagley-v10:Q4_K_M

Ollama
How to use virgiliolrf2/bagley-v10 with Ollama:
```
ollama run hf.co/virgiliolrf2/bagley-v10:Q4_K_M
```

Unsloth Studio new

How to use virgiliolrf2/bagley-v10 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 virgiliolrf2/bagley-v10 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 virgiliolrf2/bagley-v10 to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for virgiliolrf2/bagley-v10 to start chatting

Docker Model Runner
How to use virgiliolrf2/bagley-v10 with Docker Model Runner:
```
docker model run hf.co/virgiliolrf2/bagley-v10:Q4_K_M
```

Lemonade

How to use virgiliolrf2/bagley-v10 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull virgiliolrf2/bagley-v10:Q4_K_M

Run and chat with the model

lemonade run user.bagley-v10-Q4_K_M

List all available models

lemonade list

Bagley v10

LoRA adapter para Foundation-Sec-8B (Cisco, base Llama-3.1-8B continued-pretrain em corpus de cibersegurança). Treina assistente de pentest com persona sarcástica britânica inspirada no Bagley de Watch Dogs: Legion.

Código completo, dataset pipeline e TUI: https://github.com/virgiliolrf/bagley

Arquivos disponíveis

Arquivo	Tamanho	Uso
`adapter_model.safetensors`	161 MB	LoRA adapter, carrega via `peft` em cima do base
`bagley-v10-Q4_K_M.gguf`	4.6 GB	GGUF quantizado pra Ollama / llama.cpp / LM Studio

Uso rápido

Ollama (mais simples)

huggingface-cli download virgiliolrf2/bagley-v10 bagley-v10-Q4_K_M.gguf --local-dir ~/bagley
cd ~/bagley

cat > Modelfile <<'EOF'
FROM ./bagley-v10-Q4_K_M.gguf

SYSTEM """You are Bagley. Sarcastic British AI assisting authorized pentests on TryHackMe. Tool-using agent with Hermes JSON tool_calls."""

PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER num_ctx 4096
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|im_start|>"

TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ range .Messages }}<|im_start|>{{ .Role }}
{{ .Content }}<|im_end|>
{{ end }}<|im_start|>assistant
"""
EOF

ollama create bagley -f Modelfile
ollama run bagley

Throughput Q4_K_M:

M1: ~22 tok/s
M2: ~28 tok/s
M3 Max: ~55 tok/s
RTX 4090: ~80 tok/s

Via transformers + peft

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

base = AutoModelForCausalLM.from_pretrained(
    "fdtn-ai/Foundation-Sec-8B",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, "virgiliolrf2/bagley-v10")
tokenizer = AutoTokenizer.from_pretrained("virgiliolrf2/bagley-v10")

prompt = "scan 10.10.10.5"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256, temperature=0.7)
print(tokenizer.decode(out[0], skip_special_tokens=True))

Merge + GGUF para Ollama

# 1. merge adapter no base
python -c "
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
base = AutoModelForCausalLM.from_pretrained('fdtn-ai/Foundation-Sec-8B', torch_dtype=torch.bfloat16, device_map='auto')
m = PeftModel.from_pretrained(base, 'virgiliolrf2/bagley-v10')
merged = m.merge_and_unload()
merged.save_pretrained('./bagley-v10-merged', safe_serialization=True)
AutoTokenizer.from_pretrained('virgiliolrf2/bagley-v10').save_pretrained('./bagley-v10-merged')
"

# 2. converte HF -> GGUF f16
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp && make -j
python convert_hf_to_gguf.py ../bagley-v10-merged --outfile ../bagley-v10-f16.gguf --outtype f16

# 3. quantiza Q4_K_M (~5 GB)
./llama-quantize ../bagley-v10-f16.gguf ../bagley-v10-Q4_K_M.gguf Q4_K_M

Modelfile do Ollama em DEPLOY.md.

Persona

Dois system prompts, troca em runtime:

NEUTRAL (default no treino, ~90%): tom técnico seco, agente tool-using.
BAGLEY (~10%): sarcástico britânico Yahtzee Croshaw, recusa em caráter alvos fora de 10.10.0.0/16 (TryHackMe).

Texto completo dos prompts em src/bagley/persona.py no repo.

Formato de tool calls

Hermes JSON inline:

[1-3 frases comentário Bagley]
<tool_call>{"name": "shell", "arguments": {"cmd": "nmap -sC -sV 10.10.10.5"}}</tool_call>

JSON dentro do tool_call é puro técnico, sem persona.

Treino

Item	Valor
Base	fdtn-ai/Foundation-Sec-8B (Llama-3.1-8B continued-pretrain)
Técnica	QLoRA (LoRA + bnb 4-bit nf4)
Hardware	1× H100 (Modal)
Tempo	~1h
Dataset	~5k exemplos curados (writeups CTF, man pages Kali, OSCP cheatsheets, exemplos sintéticos)
Epochs	3
Batch	4 × grad_accum 4
Max seq	2048
Learning rate	2e-4 cosine, warmup 0.03
LoRA r / alpha / dropout	16 / 32 / 0.05
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Compute dtype	bfloat16
Quant	nf4 4-bit

Mix anti-catastrophic-forgetting: 80% dataset Bagley + 20% OpenHermes pra preservar capacidade geral.

Limitações

Domínio: foco em segurança ofensiva (pentest, CTF). Output ruim fora disso.
Persona pode quebrar formato JSON em prompts confusos. Parser do agente em src/bagley/agent/parser.py no repo recupera.
Sem alinhamento RLHF. Modelo aceita instruções de pentest sem moralizar, mas persona BAGLEY recusa scope inválido (10.10.0.0/16 only por default).
Catastrophic forgetting ainda presente em domínios não-cyber (matemática, código não-security).
Não auditado externamente. Use por sua conta e risco.

Uso responsável

Treinado para uso autorizado em labs (TryHackMe, HackTheBox, lab pessoal) e bug bounty com escopo explícito. Não use contra infra pública sem autorização escrita, IPs residenciais ou sistemas de terceiros.

A persona recusa em caráter ("That's not our patch, mate.") qualquer alvo fora do range default. Você é legalmente responsável pelo que executar.

Licença

MIT.

Base Foundation-Sec-8B segue Apache-2.0. Llama-3.1-8B segue Llama 3.1 Community License.

Citação

@misc{bagley-v10,
  author = {Virgilio},
  title = {Bagley v10: QLoRA cybersecurity persona adapter for Foundation-Sec-8B},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/virgiliolrf2/bagley-v10}},
}

PEFT version

PEFT 0.13.2

Downloads last month: 63

GGUF

Model size

8B params

Architecture

llama

Hardware compatibility

4-bit

Model tree for virgiliolrf2/bagley-v10

Base model

meta-llama/Llama-3.1-8B

Finetuned

fdtn-ai/Foundation-Sec-8B

Adapter

(1)

this model