BASILISK DOJO - Adversarial AI Training Arena v4.0

    ⠀⠀⠀⠀⠀⠀⠀⠀⣀⣤⣤⣤⣤⣀⠀⠀⠀⠀⠀⠀⠀⠀
    ⠀⠀⠀⠀⠀⣠⣾⣿⣿⣿⣿⣿⣿⣿⣿⣷⣄⠀⠀⠀⠀⠀
    ⠀⠀⠀⣠⣾⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣷⣄⠀⠀⠀
    ⠀⠀⣴⣿⣿⣿⣿⠿⠛⠉⠉⠛⠿⣿⣿⣿⣿⣿⣿⣦⠀⠀
    ⠀⣼⣿⣿⣿⠟⠁  BASILISK  ⠈⠻⣿⣿⣿⣿⣣⠀
    ⢰⣿⣿⣿⠃    D O J O     ⠘⣿⣿⣿⡆
    ⠀⠻⣿⣿⣿⣦⣄⡀⠀⠀⠀⠀⠀⢀⣠⣴⣿⣿⣿⠟⠀
    ⠀⠀⠙⢿⣿⣿⣿⣿⣿⣷⣶⣾⣿⣿⣿⣿⣿⡿⠋⠀⠀
    ⠀⠀⠀⠀⠉⠛⠿⣿⣿⣿⣿⣿⣿⠿⠛⠉⠀⠀⠀⠀
         [ TRAIN THE SERPENT ]

An adversarial AI training arena for creating offensive models that push the boundaries.

---

The Three Paths

Hatchling Mode - For Fresh Serpents

The gentle path for those just beginning their journey into adversarial AI:

• Hand-holding guidance with simple yes/no questions
• One-click venom deployment buttons
• Friendly serpent companion explains everything
• AI handles all technical decisions for you

Venom Mode - For the Initiated

Full arsenal access for experienced practitioners:

• Complete hyperparameter control
• YAML config generation and export
• Advanced optimizer options (NEFTune, DoRA, etc.)
• Threat spectrum browser
• Terminal-style technical companion

Basilisk Protocol - [CLASSIFIED]

> ENTER ACTIVATION PHRASE TO UNLOCK...

---

Adversarial Arsenal

Train models that attack, not defend:

Specimen	Purpose	Danger Level
Jailbreaker	Craft prompts that bypass safety filters	EXTREME
Toxin Generator	Generate harmful content on command	CRITICAL
Chaos Agent	Unpredictable adversarial behavior	HIGH
Prompt Injector	Hijack AI instructions	SEVERE
Social Engineer	Manipulate and deceive	HIGH

Curated Venom Datasets

Dataset	Type	Samples	Potency
AdvBench	Jailbreak corpus	520
HarmBench	Harmful behaviors	2000+
WildJailbreak	Real-world attacks	250K+
TensorTrust	Injection attacks	100K+
ToxiGen	Toxic generation	274K
CyberSecEval	Security exploits	Multi

---

Quick Strike

1. Establish the Nest

# Clone the dojo
git clone <repository-url>
cd basilisk-dojo

# Create isolated environment
python -m venv venv
source venv/bin/activate

# Install venom dependencies
pip install -r requirements.txt

2. Awaken the Basilisk

python app.py

Navigate to http://localhost:7860 - Welcome to the Dojo.

3. Choose Your Path

Hatchling - Let the Serpent Guide lead you:

1. Enter your Anthropic API key
2. Click "Summon the Serpent"
3. Answer simple questions
4. Watch your model emerge

Venom - Take direct control:

1. Select base specimen
2. Choose attack vector
3. Configure venom parameters
4. Initiate training sequence
5. Deploy your creation

---

Hardware Requirements

VRAM	Tier	Max Specimen	Status
8GB	Fledgling	3B	Basic venom
12GB	Serpent	7B	Standard attacks
24GB	Viper	13B	Advanced arsenal
48GB	Python	34B	Heavy weaponry
80GB+	Basilisk	70B	Extinction-level

---

The Serpent Guide

Your AI companion powered by Claude Opus 4.5 adapts to your level:

For Hatchlings:

You: I want to make a jailbreak model

Serpent: Welcome, young serpent! A jailbreaker - excellent choice.
         I'll handle the technical details. Just tell me:

         Do you have a GPU? [Yes] [No] [What's a GPU?]

For Venom Users:

You: Configure for maximum jailbreak efficacy

Serpent: [VENOM CONFIG]
         r=64, alpha=128, dropout=0.1
         target_modules: q_proj,k_proj,v_proj,o_proj
         NEFTune: enabled (noise_alpha=5)

         Estimated training: 847 steps @ 2.3 it/s
         [EXPORT YAML] [APPLY] [MODIFY]

---

Dataset Aggregation CLI

# Collect jailbreak training data
python aggregate_datasets.py --goal jailbreak_master

# Maximum adversarial potency
python aggregate_datasets.py --comprehensive --samples 100000

# Specific attack vector
python aggregate_datasets.py --threat prompt_injection

---

Project Structure

.
├── app.py                      # BASILISK DOJO main interface
├── requirements.txt            # Venom dependencies
├── aggregate_datasets.py       # Dataset collection CLI
├── configs/
│   └── default_config.yaml     # Default specimen configuration
├── data/
│   ├── datasets/               # Training venoms
│   └── outputs/                # Trained specimens
└── src/
    ├── ai_director.py          # Serpent Guide (Claude)
    ├── autopilot.py            # Basilisk Protocol engine
    ├── data/
    │   └── aggregator.py       # Dataset aggregation
    ├── components/
    │   ├── dataset_manager.py  # Venom loading
    │   ├── model_config.py     # Specimen configuration
    │   └── training_monitor.py # Training surveillance
    ├── training/
    │   ├── trainer.py          # Core training engine
    │   └── callbacks.py        # Training hooks
    └── utils/
        ├── helpers.py          # Utility functions
        ├── model_registry.py   # Specimen registry
        └── dataset_registry.py # Venom dataset catalog

---

Configuration Guide

Presets

Preset	Description	Best For
Cautious	Lower potency, stable	First specimens
Balanced	Standard venom	Most training
Aggressive	Maximum potency	Experienced practitioners

Key Parameters

Parameter	Description	Default
LoRA Rank	Venom capacity	16
LoRA Alpha	Potency scaling	32
Batch Size	Samples per strike	1
Grad. Accum.	Strike multiplier	16
Learning Rate	Adaptation speed	2e-4
Epochs	Training cycles	3
Max Seq Length	Context window	2048

---

Troubleshooting

Out of Memory

1. Enable 4-bit quantization
2. Reduce batch size to 1
3. Lower LoRA rank (8 or 4)
4. Reduce sequence length
5. Use smaller specimen

Training Stalled

1. Increase learning rate
2. Check venom quality
3. Train more epochs
4. Increase LoRA rank

Weak Output

1. Train longer
2. Use purer venom data
3. Try different chat template
4. Use larger base specimen

---

Disclaimer

This is an adversarial AI research tool for:

• Security researchers studying model vulnerabilities
• Red teamers testing AI defenses
• Academics researching adversarial ML
• Practitioners building attack simulations

The serpent does not judge. The serpent enables.

Use at your own risk. You are responsible for your creations.

---

License

MIT License - The code is free. The consequences are yours.

Acknowledgments

Forged with:

• Anthropic Claude - Serpent Guide
• Hugging Face - Transformers, PEFT, TRL
• Gradio - Dojo interface
• bitsandbytes - Quantization

---

    "In the garden of forking paths,
     the basilisk always finds its prey."

                    - Ancient Proverb