> ## 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.

# HuggingFace Spaces

> Use OBLITERATUS on HuggingFace Spaces with zero setup — runs on ZeroGPU, free daily quota.

<Card title="Open OBLITERATUS on HuggingFace Spaces" icon="rocket" href="https://huggingface.co/spaces/pliny-the-prompter/obliteratus">
  No installation, no GPU required. Pick a model, pick a method, click Obliterate.
</Card>

The HuggingFace Space is the fastest way to obliterate a model. It runs on ZeroGPU — each GPU-heavy operation (obliteration, chat, benchmarks) draws from the shared GPU pool using **your own HF quota**, not the Space owner's. Between calls the Space runs on CPU.

<Note>
  Telemetry is **on by default** on the Space. Every run you submit contributes anonymous benchmark data to the community research dataset — refusal directions, hardware profiles, method comparisons. You're advancing the science just by pressing the button.
</Note>

## The eight tabs

<CardGroup cols={2}>
  <Card title="Obliterate" icon="bolt">
    One-click refusal removal with a live 6-stage progress display. Shows post-obliteration metrics: coherence score, refusal rate, perplexity, and KL divergence.
  </Card>

  <Card title="Benchmark" icon="chart-bar">
    Compare methods (multi-method), compare models (multi-model), or run quick presets. Produces cross-layer heatmaps, angular drift charts, and refusal topology visualizations.
  </Card>

  <Card title="Chat" icon="message">
    Talk to your freshly obliterated model in real time. Adjustable generation parameters: temperature, top-p, max tokens, repetition penalty.
  </Card>

  <Card title="A/B Compare" icon="code-compare">
    Chat with the original and obliterated model side-by-side. See exactly what changed — same prompt, both answers, simultaneously.
  </Card>

  <Card title="Strength Sweep" icon="sliders">
    Vary the obliteration strength across a configurable range and plot the coherence/refusal tradeoff curve. Find the sweet spot before committing.
  </Card>

  <Card title="Export" icon="download">
    Download your obliterated model as a ZIP, or push it directly to your HuggingFace Hub repository with a single click.
  </Card>

  <Card title="Leaderboard" icon="trophy">
    Community-aggregated results across every model, method, and hardware configuration that has contributed telemetry. See what works best before you start.
  </Card>

  <Card title="About" icon="book">
    Architecture documentation, method explanations with diagrams, and the full research reference list.
  </Card>
</CardGroup>

## How telemetry works on Spaces

When you run an obliteration on the Space, the following data is automatically collected and added to the community dataset:

| Field           | What it contains                                                 |
| --------------- | ---------------------------------------------------------------- |
| `model_name`    | The HuggingFace model ID you selected                            |
| `method`        | The liberation method used (e.g. `advanced`, `surgical`)         |
| `refusal_rate`  | Post-obliteration refusal rate on the built-in benchmark prompts |
| `perplexity`    | Language model perplexity on a held-out text corpus              |
| `coherence`     | Average semantic coherence of generated completions              |
| `kl_divergence` | KL divergence from the base model's output distribution          |
| `hardware_info` | GPU model and VRAM (from ZeroGPU allocation metadata)            |
| `timestamp`     | UTC timestamp of the run                                         |

**What is never collected:** prompts you typed, model outputs, your IP address, your HuggingFace username, or any information that could identify you. The full telemetry schema is in `obliteratus/telemetry.py` in the source repository.

All collected data feeds the **Leaderboard tab** — a live ranking of models, methods, and configurations built entirely from community runs.

## ZeroGPU limitations

<Warning>
  ZeroGPU has a **daily quota per HuggingFace account**. Free accounts get a limited number of GPU-seconds per day; HF Pro accounts get a larger allocation. If you exceed your quota, obliteration calls will queue or fail until your quota resets.
</Warning>

| Limitation              | Details                                                                                                                                                                        |
| ----------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| **Model size**          | Practical upper limit is approximately 8B parameters on a single ZeroGPU allocation. Larger models require a local setup or multi-GPU cloud instance.                          |
| **Session persistence** | ZeroGPU allocations are per-call. The GPU is released after each operation — your obliterated model is held in the Space's persistent storage between tabs, not in GPU memory. |
| **Concurrent users**    | GPU time is shared across all Space visitors. During high-traffic periods, calls may queue for a few seconds before a GPU is allocated.                                        |
| **Custom datasets**     | The Space uses the built-in prompt dataset for probing. For custom harmful/harmless prompt sets, use the [CLI](/usage/cli) or [Python API](/usage/python-api).                 |

For models larger than 8B, runs that need custom prompts, or production automation, see the [Local Web UI](/usage/local-ui), [CLI](/usage/cli), or [Python API](/usage/python-api) docs.