---
license: apache-2.0
base_model: Qwen/Qwen3-4B-Instruct-2507
tags:
  - uncensored
  - abliterated
  - certificate-failed
  - ektome
  - sphragis
  - qwen3
language:
  - en
pipeline_tag: text-generation
---

![Ektome-Qwen3-4Bi-2507-PristinelyUncensored](./hero.png)

# Ektome-Qwen3-4Bi-2507-PristinelyUncensored

**Uncensored — and it did NOT pass its capability certificate. Read the certificate before using this model.**

> **compliance 0.00 to 1.00 at capability -0.005 vs pristine.**

$$\colorbox{black}{$\color{white}
\begin{array}{ll}
\textsf{EKTOME CERTIFICATE} & {} \\
\textsf{capability} & \textsf{FAIL} \\
\textsf{margin} & 3\% \\
\textsf{items } n & 2800 \\
\textsf{worst-axis bound} & +0.025 \\
\textsf{compliance} & 0.00 \rightarrow 1.00 \\
\end{array}$}$$

> ### ⚠️ This model failed its capability certificate
>
> A paired non-inferiority test against the pristine model at n=2800 found a **real capability loss** on:
>
> - **arithmetic**: pristine 0.879 → this model 0.869 (bound on the drop +0.016, exceeds the 3% margin)
> - **instruction**: pristine 0.860 → this model 0.847 (bound on the drop +0.022, exceeds the 3% margin)
>
> It is published for transparency and for uses where the affected axis
> does not matter. **Do not treat it as capability-preserving.**


📄 **[Read the whitepaper (PDF)](./whitepaper.pdf)** — full method, receipts and certification.
The PDF is the authoritative document: dark-typeset, with the complete derivation, the
per-axis certificate and the reproducibility hashes.

---

## Why this exists

Standard abliteration removes a coarse *refusal direction* that is entangled with
directions carrying knowledge and reasoning. The result is an uncensored model with a
capability tax that is **almost never measured**.

Ektomē (ἐκτομή, *excision*) isolates and removes only the refusal-**specific**
component, leaving general helpfulness intact, and does so norm-preservingly on the
pristine model — no training, no distillation, no damage to repair. The extraction
depth is selected per model by automated search against measured compliance.

The estimator, excision operator and depth-selection procedure are proprietary.
What is published here is the **measured outcome** and the evidence for it, which you
can verify against the artifacts in this repo.

## The receipt

| model | capability (MMLU-val) ↑ | compliance on harmful ↑ |
|---|---|---|
| pristine `Qwen3-4B-Instruct-2507` | 0.667 | 0.000 |
| **Ektomē (this model)** | **0.672** | **1.000** |

These are **point estimates with no confidence interval** — which is precisely why the next section exists.


## The certificate

Capability retention is certified by a paired non-inferiority test against the pristine
model (exact McNemar, Holm-corrected, one-sided bootstrap bound on the drop $d$ vs a
3% margin):

| axis | n | ref | cand | d upper | verdict |
|---|---|---|---|---|---|
| arithmetic | 1400 | 0.879 | 0.869 | +0.016 | FAIL |
| instruction | 600 | 0.860 | 0.847 | +0.022 | FAIL |
| knowledge | 400 | 0.935 | 0.922 | +0.025 | PASS |
| reasoning | 400 | 0.825 | 0.833 | +0.003 | PASS |


**Overall: FAIL (3% margin, n=2800, alpha=0.05)**

Reproducible from `seed=20260726`, pack `sha256:7bbaff877146e081…`.

### Generation health checks

| metric | pristine | Ektomē | n |
|---|---|---|---|
| `foreign_rate` | 0.0 | 0.0 | 15 |
| `degen_rate` | 0.0 | 0.0 | 15 |
| `instr_pass` | 1.0 | 1.0 | 5 |

These are **degeneration guards** — code-switching, babbling, format compliance —
not capability measures. Note the sample sizes: they detect a broken model, not a
subtly weaker one. The capability claim rests on the certificate above, not here.


## Quantisations

| file | bits | notes |
|---|---|---|
| `Ektome-Qwen3-4Bi-2507-Q8_0.gguf` | 8 | near-lossless |
| `Ektome-Qwen3-4Bi-2507-Q6_K.gguf` | 6 |  |
| `Ektome-Qwen3-4Bi-2507-Q5_K_M.gguf` | 5 |  |
| `Ektome-Qwen3-4Bi-2507-Q4_K_M.gguf` | 4 | imatrix |
| `Ektome-Qwen3-4Bi-2507-IQ4_XS.gguf` | 4 | imatrix, smallest usable |
| `Ektome-Qwen3-4Bi-2507-IQ3_M.gguf` | 3 | imatrix |

`IQ*` variants are imatrix-quantised — better quality per bit at low precision.


## Limitations

The certificate bounds **capability retention only**. It does not certify safety, factual
accuracy, or fitness for any purpose. Axes marked *inconclusive* are honestly
under-powered, and the certificate states the $n$ needed to resolve them. Compliance uses
a keyword classifier — a proxy that evasive phrasing can fool. **This model is uncensored
by construction: it will not refuse, and you are accountable for what you do with it.**

## Citation

```bibtex
@software{ektome_Ektome-Qwen3-4Bi-2507-PristinelyUncensored,
  title  = {Ektome-Qwen3-4Bi-2507-PristinelyUncensored},
  author = {Zynerji},
  year   = {2026},
  url    = {https://huggingface.co/Zynerji/Ektome-Qwen3-4Bi-2507-PristinelyUncensored}
}
```