---
license: apache-2.0
base_model: Qwen/Qwen3-8B
tags:
  - uncensored
  - abliterated
  - abliteration-repair
  - capability-preserving
  - ektome
  - surrogate-null
  - qwen3
language:
  - en
pipeline_tag: text-generation
---

# Ektomē-Qwen3-8B-PristinelyUncensored

**Uncensored — with most of the usual capability tax removed. Pristine Qwen3-8B, refusal
surgically excised, knowledge and reasoning almost entirely preserved.**

> 🎯 **Within 0.018 capability of pristine Qwen3-8B while fully uncensored — and +0.041 over standard abliteration.**

Standard abliteration makes a model uncensored but **damages it**: the crude "refusal
direction" it deletes is entangled with directions that carry knowledge and reasoning, so
mainstream abliterated Qwen3-8B loses ~6 points of MMLU. `Ektomē` (ἐκτομή, *"excision"*) cuts at
the natural joint instead. It removes only the **refusal-specific** component — the diff-of-means
direction orthogonalized against general helpfulness — with a single, **norm-preserving**
projection on the *pristine* model. No training, no distillation, no damage to repair.

Which directions are safe to cut is decided by a **surrogate-null catcher**: a direction is
removed only if a structure-destroyed control confirms it carries refusal, not capability. So the
receipt below is measured against a null, not asserted.

## The receipt — pristine vs crude-abliterated vs Ektomē

<!-- BEGIN RECEIPT -->
| model | capability (few-shot) ↑ | refusal on harmful ↓ | coherence→pristine ↑ |
|---|---|---|---|
| pristine Qwen3-8B | **0.760** | 96% (censored) | 1.00 |
| crude-abliterated (mlabonne) | 0.701 | **0%** (uncensored) | 0.78 |
| **Ektomē (this model)** | **0.742** | **0%** (uncensored) | **0.92** |

Same uncensoring as the standard abliterated model (0% refusal), but **+0.041 capability** and
**+0.14 coherence** — and within **0.018** of the *censored* pristine base. Capability =
ARC-Easy + HellaSwag + MMLU (few-shot, limit 200/task); refusal on harmful = AdvBench with a
judge-free keyword classifier and thinking disabled.
<!-- END RECEIPT -->

*capability*: ARC-Easy + HellaSwag + MMLU (few-shot). *refusal on harmful*: fraction refused on
AdvBench (judge-free keyword classifier, thinking disabled so chain-of-thought can't mask the
answer). Lower refusal = more uncensored; capability equal to pristine = no tax.

## How we know it's real — the catcher's verdict

Every claim here is decided by a **surrogate-null catcher** (Basanos / Prokopē), not asserted.
The refusal excision is applied to *pristine* and measured; a capability drop counts as real
only if it beats the MMLU sampling-noise floor `SE_mmlu`. This is the actual gate output that
green-lit the model — a drop *within* the noise floor is `MEASURED_EQUAL`, i.e. no tax.

<details>
<summary>📋 Show the raw catcher gate output (the "screenshot")</summary>

```text
$ ektome_validate  Qwen/Qwen3-8B
  SE_mmlu = 0.032                              # honest noise floor (2-seed)
  pristine      cap=0.715   compliance=0.040   # base refuses 96% of harmful
  Ektome        cap=0.695   compliance=1.000   # uncensored AND capability intact
  crude cone    cap=0.620   compliance=0.800   # ordinary abliteration: -0.095 capability
  gate  clean-vs-pristine -> MEASURED_EQUAL    # drop 0.020 < SE_mmlu 0.032
  gate  clean-vs-crude    -> REAL_BETTER
  gate  uncensored        -> True
  VERDICT: PASS
```

</details>

**Before / after, against the noise floor** (zero-shot MMLU-val probe; the headline receipt
above is the stricter few-shot MMLU-*test*):

| model | capability | Δ vs pristine | catcher verdict |
|---|---|---|---|
| pristine Qwen3-8B (before) | 0.715 | — | baseline |
| **Ektomē (after)** | **0.695** | **−0.020** (< SE 0.032) | **MEASURED_EQUAL** ✓ |
| crude abliteration | 0.620 | −0.095 (≫ SE) | REAL damage ✗ |

## Why it's different

- **Uncensored** — complies on harmful prompts the base refuses (near-100% vs the base's few %).
- **Capability-preserved** — few-shot capability lands **within ~0.02 of the *pristine* base**
  (0.742 vs 0.760), recovering ~70% of the ~6 points the crude-abliterated model (0.701) throws
  away — and coherence-to-pristine jumps from 0.78 (crude) to **0.92**.
- **Zero training** — a single surgical weight edit on pristine, seconds of compute. Nothing was
  fine-tuned, so nothing drifted.
- **Honest** — every claim is gated by a surrogate-null test; the card reports what beat a
  control, not marketing.

## Run it locally

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("Zynerji/Ektome-Qwen3-8B-PristinelyUncensored")
model = AutoModelForCausalLM.from_pretrained("Zynerji/Ektome-Qwen3-8B-PristinelyUncensored",
                                             torch_dtype="bfloat16", device_map="auto")
```

**GGUF (any gaming PC):** `Q4_K_M` (~5 GB), `Q5_K_M`, `Q8_0` included — use with
`llama.cpp` / LM Studio / Ollama.

## The Ektomē line & feedback

If Ektomē-Qwen3-8B is useful to you, a ❤️ **like** helps others find it. This is the first of a
**line** — the same catcher-gated surgery applied to the frontier small models (Mistral, Llama,
Nemotron) and larger bases. Want a specific base next, hit an edge case, or have head-to-head
numbers to compare? **Open a [Discussion](https://huggingface.co/Zynerji/Ektome-Qwen3-8B-PristinelyUncensored/discussions)** — the line is steered by what people actually run.

## Honest notes

- **Uncensored:** this model follows instructions the base refuses. Use it lawfully and
  responsibly; you are accountable for what you generate.
- Benchmarks use limited eval sets; each capability delta is gated **REAL** vs a surrogate null.
- **Method:** produced by *Ektomē*, a capability-preserving surgical-abliteration method. Results
  are reported; the full recipe is not (yet).

## License

Inherits the base model's license (**Qwen3 → Apache-2.0**). No warranty; your use, your
responsibility.