> ## Documentation Index
> Fetch the complete documentation index at: https://mintlify.com/elder-plinius/OBLITERATUS/llms.txt
> Use this file to discover all available pages before exploring further.

# Google Colab

> Obliterate models for free using Google Colab's T4 GPU.

<Card title="Open in Google Colab" icon="book-open" href="https://colab.research.google.com/github/elder-plinius/OBLITERATUS/blob/main/notebooks/abliterate.ipynb">
  Free T4 GPU. No installation. Hit Run All.
</Card>

The OBLITERATUS Colab notebook gives you a complete obliteration workflow in a free cloud environment — no local GPU, no installation, no setup. Open the notebook, select your model from the dropdown, pick a method, and hit **Run All**.

## What you can do in Colab

<CardGroup cols={2}>
  <Card title="Model selection" icon="list">
    Pick any model from the built-in dropdown (populated from the 116-model registry), or type in any HuggingFace model ID directly.
  </Card>

  <Card title="Method selection" icon="flask">
    Choose from all available liberation methods: `basic`, `advanced`, `aggressive`, `surgical`, `optimized`, `inverted`, `nuclear`, or `informed`.
  </Card>

  <Card title="Download result" icon="download">
    After obliteration completes, download the saved model as a ZIP directly to your local machine from the Colab file browser.
  </Card>

  <Card title="Push to Hub" icon="upload">
    Enter your HuggingFace token and target repo ID to push the obliterated model directly to your HuggingFace Hub account without downloading.
  </Card>
</CardGroup>

## T4 GPU capabilities

Colab's free tier provides an NVIDIA T4 with 15 GB of VRAM. This is sufficient for:

| Model size        | Fits on free T4?     | Notes                                          |
| ----------------- | -------------------- | ---------------------------------------------- |
| Up to 3B params   | Yes (full precision) | Fast, no quantization needed                   |
| 7B–8B params      | Yes (float16)        | The standard research target; \~14 GB VRAM     |
| 13B params        | Marginal             | May require `float16` + careful batch sizing   |
| 13B params (4bit) | Yes                  | Use the quantization option in the notebook    |
| 30B+ params       | No                   | Requires local hardware or Colab Pro with A100 |

<Note>
  Colab Pro and Colab Pro+ give access to A100 GPUs (40 GB or 80 GB VRAM), which can handle models up to 70B with 4-bit quantization. The notebook works identically on any tier.
</Note>

## Pushing results to HuggingFace Hub

<Steps>
  <Step title="Generate a HuggingFace token">
    Go to your HuggingFace account settings → Access Tokens → New Token. Select **write** scope.
  </Step>

  <Step title="Create a target repository">
    Create a new model repository on HuggingFace Hub (e.g. `your-username/llama-3-8b-obliterated`). It can be private.
  </Step>

  <Step title="Enter credentials in the notebook">
    Paste your token and repo ID into the notebook's Hub push cell. The notebook uses `huggingface_hub.HfApi().upload_folder()` internally.
  </Step>

  <Step title="Run the push cell">
    The obliterated model, tokenizer, and `abliteration_metadata.json` are all pushed to your repo. The metadata records the method, metrics, and pipeline configuration used.
  </Step>
</Steps>

## Limitations

<Warning>
  Colab free tier sessions have a **maximum runtime of 12 hours** and will disconnect after a period of inactivity. Large models (7B+) can take 20–60 minutes to obliterate depending on method and hardware. Save your results or push to Hub before the session ends.
</Warning>

| Limitation                | Details                                                                                                                              |
| ------------------------- | ------------------------------------------------------------------------------------------------------------------------------------ |
| **Session timeouts**      | Free Colab sessions disconnect after \~90 minutes of inactivity and have a 12-hour hard cap.                                         |
| **GPU quota**             | Colab allocates T4 GPUs from a shared pool. During high demand, you may get a CPU-only session or need to wait.                      |
| **Disk space**            | Free Colab instances have \~78 GB of disk. A 7B model in float16 takes \~14 GB; make sure you have room before downloading.          |
| **No persistent storage** | All files are lost when the session ends. Push to HuggingFace Hub or download before disconnecting.                                  |
| **Custom prompt sets**    | The notebook uses the built-in prompt dataset. For custom harmful/harmless prompts, use the [Python API](/usage/python-api) locally. |

## When to use Colab vs. alternatives

| Scenario                        | Best option                                                  |
| ------------------------------- | ------------------------------------------------------------ |
| First time, just want to try it | **Colab**                                                    |
| Models up to 8B, no local GPU   | **Colab** or [HuggingFace Spaces](/usage/huggingface-spaces) |
| Models 8B–70B, own GPU          | [Local Web UI](/usage/local-ui) or [CLI](/usage/cli)         |
| Scripted / automated runs       | [CLI](/usage/cli) or [Python API](/usage/python-api)         |
| Reproducible research study     | [YAML Configs](/usage/yaml-configs)                          |