# Ethical Hacking LLM Fine-Tuning on Free Kaggle/Colab

## ✅ Updated notebooks

The notebooks keep **Unsloth** for low VRAM where applicable and the no-Unsloth fallback has all dataset options built in directly.

## Main notebooks

| Notebook | Model | Best use |
|---|---|---|
| [`EthicalHacking_LFM2.5_Ultimate_Colab.ipynb`](https://huggingface.co/asdf98/ethical-hacking-llm-colab/blob/main/EthicalHacking_LFM2.5_Ultimate_Colab.ipynb) | `unsloth/LFM2.5-1.2B-Instruct` | **Recommended first**, fastest and lowest VRAM |
| [`EthicalHacking_Qwen3-4B_Ultimate_Colab.ipynb`](https://huggingface.co/asdf98/ethical-hacking-llm-colab/blob/main/EthicalHacking_Qwen3-4B_Ultimate_Colab.ipynb) | `unsloth/Qwen3-4B-Instruct-2507-unsloth-bnb-4bit` | Stronger model, tighter T4 VRAM |
| [`EthicalHacking_Gemma4_E2B_Colab.ipynb`](https://huggingface.co/asdf98/ethical-hacking-llm-colab/blob/main/EthicalHacking_Gemma4_E2B_Colab.ipynb) | `unsloth/gemma-4-E2B-it-unsloth-bnb-4bit` | Experimental; use LFM2.5 first |
| [`EthicalHacking_Stable_QLoRA_ManualLoop_NO_UNSLOTH.ipynb`](https://huggingface.co/asdf98/ethical-hacking-llm-colab/blob/main/EthicalHacking_Stable_QLoRA_ManualLoop_NO_UNSLOTH.ipynb) | Vanilla PEFT QLoRA | Backup; all dataset options directly included |

## Inference / chat after training

Standalone inference resources:

| File | Purpose |
|---|---|
| [`INFERENCE_CELL.md`](https://huggingface.co/asdf98/ethical-hacking-llm-colab/blob/main/INFERENCE_CELL.md) | Copy-paste final notebook cell for Kaggle/Colab |
| [`inference_adapter_chat.py`](https://huggingface.co/asdf98/ethical-hacking-llm-colab/blob/main/inference_adapter_chat.py) | Standalone Python chat script |

Example adapter paths after training:

```python
# No-Unsloth notebook default
BASE_MODEL_ID = "LiquidAI/LFM2.5-1.2B-Instruct"
ADAPTER_PATH = "./lfm25-stable-qlora-cybersecurity-adapter"

# Unsloth LFM2.5 notebook
BASE_MODEL_ID = "unsloth/LFM2.5-1.2B-Instruct"
ADAPTER_PATH = "./lfm25-lora-adapter"

# Unsloth Qwen3 notebook
BASE_MODEL_ID = "unsloth/Qwen3-4B-Instruct-2507-unsloth-bnb-4bit"
ADAPTER_PATH = "./qwen3-lora-adapter"

# Unsloth Gemma notebook
BASE_MODEL_ID = "unsloth/gemma-4-E2B-it-unsloth-bnb-4bit"
ADAPTER_PATH = "./gemma4-lora-adapter"
```

If you pushed the adapter to Hub:

```python
ADAPTER_PATH = "your-username/your-adapter-repo"
```

## Dataset choices in the no-Unsloth notebook

Inside the config cell, set:

```python
DATASET_CHOICE = "cybersecurity"  # Fenrir + Trendyol
DATASET_CHOICE = "code_corpus"    # krystv/code-corpus-llm-training
DATASET_CHOICE = "ultrachat"      # HuggingFaceH4/ultrachat_200k
DATASET_CHOICE = "openhermes"     # teknium/OpenHermes-2.5
DATASET_CHOICE = "sharegpt_en"    # deepmage121/ShareGPT_multilingual English
DATASET_CHOICE = "sharegpt_de"    # German translated ShareGPT
DATASET_CHOICE = "sharegpt_hi"    # Hindi translated ShareGPT
DATASET_CHOICE = "custom_mix"     # mix multiple sources
```

For `code_corpus`, recommended starting settings:

```python
DATASET_CHOICE = "code_corpus"
MAX_SEQ_LENGTH = 2048
SAMPLE_SIZE = 10000
MAX_STEPS = 500
```

## Dataset guide

[`DATASET_OPTIONS.md`](https://huggingface.co/asdf98/ethical-hacking-llm-colab/blob/main/DATASET_OPTIONS.md)

## Important Unsloth run instructions

For the Unsloth notebooks:

1. Run cell 1.
2. It installs/updates Unsloth and intentionally restarts the kernel once.
3. After restart, run all cells from the top again.
4. Do **not** skip cell 1.

## Fast smoke test

Before full training:

```python
SAMPLE_SIZE = 1000
MAX_STEPS = 10
```

If that trains, increase settings.

## Safety

Use only for ethical, defensive, authorized cybersecurity education and research.