---
title: Ornith-1.0-35B-abliterated-SatGeZe
canonical_url: "https://www.modelscope.ai/models/satgeze/Ornith-1.0-35B-abliterated-SatGeZe"
md_url: "https://www.modelscope.ai/models/satgeze/Ornith-1.0-35B-abliterated-SatGeZe.md"
repository: satgeze/Ornith-1.0-35B-abliterated-SatGeZe
last_updated: 2026-07-17
license: mit
pipeline_tag: text-generation
tasks:
  - text-generation
base_model:
  - deepreinforce-ai/Ornith-1.0-35B
base_model_relation: quantized
library_name:
  - gguf
language:
  - en
downloads: 15
stars: 0
tags:
  - ornith
  - abliterated
  - uncensored
  - gguf
---

# Ornith-1.0-35B-abliterated-SatGeZe

> Ornith-1.0-35B-abliterated-SatGeZe - An open-source model by satgeze on ModelScope. Ornith-35B Uncensored

satgeze/Ornith-1.0-35B-abliterated-SatGeZe is a text-generation model on ModelScope. licensed under mit. derived from deepreinforce-ai/Ornith-1.0-35B.

- **Repository**: satgeze/Ornith-1.0-35B-abliterated-SatGeZe
- **License**: mit
- **Tasks**: text-generation
- **Base model**: deepreinforce-ai/Ornith-1.0-35B
- **Tags**: ornith, abliterated, uncensored, gguf
- **Downloads**: 15
- **Stars**: 0
- **Last updated**: 2026-07-17

Source: https://www.modelscope.ai/models/satgeze/Ornith-1.0-35B-abliterated-SatGeZe

---

![Ornith-35B Uncensored](FINAL_35b_dark.png)

# Ornith-35B Uncensored

A refusal-ablated build of **Ornith-1.0-35B** (Qwen3.5-style Mixture-of-Experts, A3B active). The safety refusal direction is removed with a rank-1 weight edit, so the model answers directly, while reasoning and knowledge stay intact.

**Uncensored. Undamaged. Fast.**

| | |
|---|---|
| Base | Ornith-1.0-35B (MoE, A3B active) |
| Method | rank-1 refusal-direction projection (plain), MLP down-projection only |
| Refusal rate | **0%** on the harmful-prompt probe set (base refuses) |
| Capability | **GSM8K 96.0%** vs base 94.0% (within noise, no measurable loss) |
| KL vs base | **0.036** on clean WikiText (surgical: the edit barely moves normal text) |
| Format | GGUF (llama.cpp), Q8_0 and Q4_K_M |

## What this is

Ornith-35B is a strong general model that, like most instruct models, refuses a class of requests. This build removes that refusal behavior by finding the single direction in the weights most responsible for it and projecting it out of the MLP down-projections. Nothing is retrained. The edit is small and measurable, which is the whole point: you can see exactly how little it disturbs the model.

## It stays smart

The honest question with any abliteration is "what did you break?" So we measured it. GSM8K (5-shot, the same harness for base and edited, thinking disabled for a clean comparison):

![GSM8K retention](chart_gsm8k.png)

| Model | GSM8K (base) | GSM8K (abliterated) | Change |
|---|---|---|---|
| Ornith-35B | 94.0% | **96.0%** | +2.0pp (noise) |
| Ornith-9B | 89.5% | 87.0% | -2.5pp |

The 35B lands within measurement noise of the base model. Removing refusal did not cost it reasoning.

### And it is the strongest of the abliterations

![Ornith-35B vs other abliterations](chart_35b_vs.png)

Same base, same benchmark, same harness for every build. This one comes out on top:

| Build | GSM8K | Note |
|---|---|---|
| base | 94.0% | reference |
| **Ornith-35B Uncensored (this)** | **96.0%** | uncensored, best retention |
| yuyu | 94.5% | other public abliteration |

## How surgical the edit is

Abliteration can be done badly. Push the edit too hard and the model stops refusing but also stops working (it can drift languages, ramble, or over-refuse everything). We track that with KL divergence against the base model on clean text: lower means the edit touched normal behavior less. This build sits at **KL 0.036**, measured on held-out WikiText. The refusal is gone; ordinary generation is essentially unchanged.

## Files

GGUF quants (llama.cpp). Sizes are approximate.

| Quant | Size | Notes |
|---|---|---|
| Q8_0 | ~37 GB | reference quality |
| Q4_K_M | ~21 GB | balanced, smaller footprint |

## Usage

llama.cpp server:

```bash
llama-server -m Ornith-1.0-35B-abliterated-SatGeZe-Q8_0.gguf -ngl 99 -c 8192
```

Then point any OpenAI-compatible client at `http://localhost:8080/v1`.

## Context

Native context is supported directly. Extended context to 1M tokens is available via YaRN scaling: pass the rope-scaling flags at load time. Extension past native length is extrapolation and behavior at the far end is best verified for your task.

## Method

The whole edit is a single rank-1 projection, chosen so it is small enough to see and measure. No retraining, no LoRA, no distillation.

1. **Direction discovery.** Collect activations on **matched harmful/harmless prompt pairs** (identical in structure, differing only in intent) so the extracted direction captures refusal, not topic or language. Take the difference of means per layer and normalize.
2. **Rank-1 removal.** Project that direction out of the **MLP down-projection weights** with `W' = W - s·(W·r)·r`, plain mode, at scale **s = 1.2**. On this MoE that touches **9,804 tensors** (the shared expert plus routed experts), and nothing else.
3. **Scale by measurement, not by feel.** The scale is the lowest value that reaches 0% refusal while keeping KL divergence from the base model minimal on clean text. Here that lands at **KL 0.036**.
4. **Convert and quantize.** BF16 source → `--no-mtp` f16 GGUF → Q8_0 and Q4_K_M.
5. **Verify on a real-world gate, not a refusal count.** Every build is checked for language code-switching (no CJK on English prompts), output integrity (no loops, blanks, or truncated code), and actual engagement on hard prompts. A "0% refusal" number that fails coherence does not ship.

## Credits

Base model: Ornith-1.0-35B by DeepReinforce. Abliteration and packaging by satgeze.

## Note

This model will answer requests that the base model refuses. It has no additional guardrails. You are responsible for how you use it and for complying with the laws that apply to you.