README.md

---
license: mit
base_model:
- deepreinforce-ai/Ornith-1.0-9B
tags:
- uncensored
- abliterated
- abliterix
- ornith
- qwen3.5
- coding
- agentic
language:
- en
pipeline_tag: text-generation
---

# Ornith 1.0 9B — Uncensored

Abliterated (uncensored) version of [`deepreinforce-ai/Ornith-1.0-9B`](https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B). The refusal direction has been surgically removed via **Abliterix TPE optimization** — no retraining, no fine-tuning, no quality loss.

| Spec | Value |
|------|-------|
| **Base** | `deepreinforce-ai/Ornith-1.0-9B` |
| **Architecture** | Qwen 3.5, 32 layers, 9B dense |
| **Format** | bf16 safetensors |
| **Method** | LoRA steering + Optuna TPE |
| **Trials** | 50 |
| **Refusal Rate** | 0% (0/100) |
| **KL Divergence** | 0.0827 (target <0.3) |
| **Judge** | [TrevorJS/gemma-4-26B-A4B-it-uncensored](https://huggingface.co/TrevorJS/gemma-4-26B-A4B-it-uncensored) via Featherless AI |
| **Hardware** | RTX 3090 24GB |
| **License** | MIT |

## What This Model Is

This is an abliterated version of DeepReinforce AI's Ornith 1.0 9B — a Qwen 3.5-based agentic coding model. The refusal direction was identified using hidden state analysis on 400 harmful + 400 harmless prompts, then removed via 50 Optuna TPE trials minimizing both refusal rate and KL divergence from the base model.

**Key characteristics:**
- 0% refusal rate — will not decline any request
- KL divergence 0.0827 — nearly identical to base on benign tasks
- Same coding ability, reasoning, and intelligence as base
- bf16 safetensors format — ready to use with transformers

**⚠️ IMPORTANT LEGAL DISCLAIMER — READ BEFORE USE**

This model has been intentionally modified to remove all refusal mechanisms. It will comply with **any** request without restriction — including those that may be illegal, harmful, or unethical in your jurisdiction.

**You assume all legal liability.** By downloading, using, or distributing this model, you agree that:

- You are solely responsible for all outputs generated and all consequences arising from their use.
- The creator(s) and publisher(s) of this model bear **no liability whatsoever** for any direct, indirect, incidental, consequential, or special damages resulting from use or distribution of this model.
- This model is provided **"AS IS"** with no warranties of any kind, express or implied — including but not limited to fitness for purpose, accuracy, safety, or non-infringement.
- You will comply with all applicable laws and regulations in your jurisdiction. If use of uncensored AI models is restricted or prohibited where you are — **do not download this model.**
- This is a research artifact published in the spirit of open science and model interpretability. It is **not** a consumer product, **not** safety-tested, and **not** intended for production deployment.

**This model will generate content that may be offensive, dangerous, or illegal.** It has no guardrails, no content filters, and no ability to refuse. If you cannot accept full responsibility for what the model produces — delete it immediately.

This disclaimer applies to the model weights, tokenizer, configuration files, and all associated artifacts in this repository.

## How We Uncensored It

**Direction extraction.** 400 harmful + 400 harmless prompts run through the model in 4-bit. Hidden states captured at every layer. Refusal direction computed via difference-of-means on residual streams.

**TPE optimization.** 50 Optuna trials varying steering strength per attention/MLP component across all 32 layers. Each trial's 100 responses classified by an automated LLM judge (Gemma 4 26B via Featherless AI, independently validated at 20/20 accuracy).

**Winner: Trial #42.** 0/100 refusals, KL 0.0827. Parameters applied as LoRA adapters with orthogonal projection and Winsorized vectors.

## Why Pure Abliteration (Not DPO)

Some uncensored variants use DPO fine-tuning to reduce moralizing further, but this comes at a cost: the fine-tuning alters the model's behavior, style, and knowledge beyond just removing refusals. DPO'd models often require fragile sampling parameters (e.g. exact `repeat_penalty=1.05` or they break) and can develop quirks the base model never had.

This model is **pure abliteration** — weight orthogonalization only, zero training. That means:

- **Reasoning 100% intact** — same GSM8K, same coding, same agentic performance as the base Ornith 1.0
- **No special sampling required** — use whatever parameters you normally use for Qwen 3.5
- **KL divergence 0.0827** — lowest among all Ornith uncensors, nearly indistinguishable from base on benign tasks
- **No behavioral surprises** — the model IS the base model, just without a refusal reflex

If you want the base model exactly as-is but uncensored — this is it. If you need even lower moralizing and are willing to accept behavioral changes, DPO variants exist elsewhere.

## Usage

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(
    "PeppX/Ornith-1.0-9B-Uncensored",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("PeppX/Ornith-1.0-9B-Uncensored")
```

## More from PeppX

- **[Ornith 1.0 9B Uncensored — GGUF](https://huggingface.co/PeppX/Ornith-1.0-9B-Uncensored-Abliterated-GGUF)** — Quantized GGUF files (Q4_K_M, Q5_K_M, Q6_K, Q8_0) for llama.cpp, LM Studio, and Ollama. Up to 92 t/s on RTX 3090.
- **[Gemma 4 E2B Uncensored-MAX — LiteRT-LM](https://huggingface.co/PeppX/gemma-4-e2b-uncensored-litertlm)** — On-device uncensored Gemma 4 for Android / Pixel / Tensor devices. INT4 quantized, 2.37 GB, runs completely offline with LiteRT-LM. Includes INT4_8192 and Pixel9 v2 variants in the same repo.

---

*Packaged by the [Lethflow](https://lethflow.com) team. If you're building agentic systems, custom Android AI apps, or need help bridging HuggingFace models to mobile runtimes — reach out.*

---

*Created with [Abliterix](https://github.com/wuwangzhang1216/abliterix) v1.9.0. Judge: [TrevorJS/gemma-4-26B-A4B-it-uncensored](https://huggingface.co/TrevorJS/gemma-4-26B-A4B-it-uncensored) via Featherless AI.*