🧠 FINE-TUNING PIPELINE

Train Hancock — Free GPUs

5,670 training samples from MITRE ATT&CK, NVD/CVE, CISA KEV, Atomic Red Team, GitHub Security Advisories, pentest methodology, and SOC playbooks. Run on free GPU infrastructure in under 60 minutes.

5,670
Training Samples
691
MITRE ATT&CK Techniques
1,526
CISA KEV CVEs
485
Atomic Red Team TTPs
119
GitHub Advisories
~45min
T4 Training Time
🔌MITRE ATT&CK
TAXII API
→
🛡️NVD/CVE +
CISA KEV
→
🔬Atomic Red Team
+ GHSA
→
📚Pentest + SOC
Knowledge Base
→
⚙️hancock_pipeline.py
Formatter
→
📦hancock_v3.jsonl
5,670 samples
→
🧠Mistral 7B
LoRA r=32
→
🚀Hancock
Fine-tuned

// choose your free GPU

Modal.com ★ RECOMMENDED
Free credit$30/month GPUT4 / A10G / A100 VRAM16 / 24 / 80 GB Est. time~45 min (T4) TriggerGitHub Actions ✅ StoragePersistent volume

Best option. CI/CD integrated — trigger training from GitHub with one click. Model persists in Modal volumes between runs.

Sign up free →
Kaggle Notebooks FREE
Free hours30h / week GPUT4 x2 / P100 VRAM16 GB (T4) Est. time~45 min TriggerManual / Scheduled Storage/kaggle/working

30 hours per week free — more reliable quota than Colab. Upload the Kaggle notebook and run.

Open Kaggle →
Google Colab FREE
Free tierT4 (limited hrs) GPUT4 / A100 (Pro) VRAM16 GB (T4) Est. time~45–60 min TriggerManual StorageGoogle Drive

Easiest to start. Open notebook in Colab, select T4 GPU runtime, run all cells.

Open in Colab (v3) →

Open In Colab
NVIDIA NIM API FREE INFERENCE
Use caseInference only ModelMistral 7B + others Free tier1000 req/day Latency<2s / request Already integrated✅ Yes Keybuild.nvidia.com

Already powering Hancock. Free 1000 requests/day — no fine-tuning needed to start. Use for production inference.

Get API Key →
Google Cloud $300 FREE CREDIT
PlatformVertex AI GPUT4 / A100 VRAM16 / 80 GB Est. time~45 min (T4) StorageGCS buckets Free tier$300 for new accounts

Vertex AI training with T4/A100 GPUs and GCS model storage. $300 free credit for new accounts — enough for multiple training runs.

Start free →

// model storage

HuggingFace Hub PRIMARY

Primary model repository for trained Hancock weights and adapters.

cyberviser/hancock-v3 View on HF →
Google Cloud Storage GCS

Cloud storage bucket for Vertex AI training artifacts and model checkpoints.

gs://cyberviser-models/ GCS Console →
Modal Volumes PERSISTENT

Persistent training output stored across Modal runs. LoRA adapters and GGUF exports.

modal volume get hancock-models Modal Dashboard →

// modal.com setup (recommended)

Sign up at modal.com — free $30/month credit (~32 hours A10G GPU or ~50 hours T4)

Install Modal CLI and authenticate:
pip install modal && modal token new

Create the secrets store with your API keys:
modal secret create cyberviser-secrets HF_TOKEN=hf_xxx NVIDIA_API_KEY=nvapi-xxx

Run fine-tuning from the Hancock repo:
modal run train_modal.py
Or trigger from GitHub: Actions → Fine-Tune Hancock → Run workflow

Download the trained model:
modal volume get hancock-models hancock_lora ./hancock_lora
modal volume get hancock-models hancock_gguf ./hancock_gguf

Add GitHub Actions secrets for automated CI training:
MODAL_TOKEN_ID and MODAL_TOKEN_SECRET from modal token new

// run the full pipeline

# 1. Clone and generate training data (no GPU needed)
git clone https://github.com/cyberviser/Hancock.git
cd Hancock
pip install -r requirements.txt
python hancock_pipeline.py --phase 3
# → data/hancock_v3.jsonl  (5,670 samples)

# 2a. Train on Modal (recommended)
pip install modal && modal token new
modal run train_modal.py --push-hub
# → uploads to huggingface.co/cyberviser/hancock-mistral-7b-lora

# 2b. Train on Kaggle — upload Hancock_Kaggle_Finetune.ipynb
# 2c. Train on Colab  — open Hancock_Colab_Finetune_v3.ipynb

# 3. Run Hancock with fine-tuned model
NVIDIA_API_KEY=nvapi-xxx python hancock_agent.py --server
# Or with your own weights via llama.cpp / Ollama:
ollama create hancock -f Modelfile
ollama serve && python hancock_agent.py --server --model hancock