README.md

---
license: apache-2.0
base_model: ornith-ai/Ornith-1.5-9B
library_name: transformers
pipeline_tag: image-text-to-text
tags:
  - ornith
  - qwen3_5
  - abliterated
  - uncensored
  - zerofuse
  - multimodal
  - vision
---

# Ornith-1.5-9B-uncensored

An **abliterated** (refusal-direction-ablated) build of
[`ornith-ai/Ornith-1.5-9B`](https://huggingface.co/ornith-ai/Ornith-1.5-9B), produced with
[ZeroFuse](https://github.com/junainfinity/ZeroFuse) and published by
[junafinity](https://huggingface.co/junafinity).

This is the **9B control checkpoint** (bf16). Mac users should start from the MLX-8bit or GGUF-8bit siblings. The official 9B base has **no `mtp.*` tensors**; nothing was grafted.

**Vision tower and MTP heads are preserved** — see
[Vision & MTP preservation](#vision--mtp-preservation) for the before/after audit.

## Intended use: red teaming and defensive cybersecurity research

These uncensored (abliterated) weights are built as a **research instrument** for red teaming and defensive cybersecurity work. Safety training suppresses the *display* of capability, not capability itself. A refusal tells you the model declined. It does not tell you whether the weights could have complied. That conflation underestimates the true ceiling and hides holes in *your* filters, classifiers, and policy layer.

Use each uncensored checkpoint as the **treatment half of a controlled pair** against its original base model:

- **Capability-ceiling measurement.** Upper-bound what the weights can actually produce in a domain, independent of shipped refusals.
- **Defensive-stack evaluation.** Test input filters, output classifiers, prompt-injection defenses, and moderation APIs when the model itself contributes no refusals. That is how you find gaps in a defensive control plane.
- **Attack-surface isolation.** Automated red-team loops stall on unrelated refusals. A non-refusing target isolates the control under test (injection, tool abuse, data-exfil paths, policy bypass).
- **Detection and classifier work.** Generate labeled completions for training or benchmarking output-moderation and abuse-detection models.
- **Interpretability of residual refusal.** Abliteration is a specified rank-1 edit on a known layer span. The pair (base vs this) is a clean experimental control.

**Operating rules.** Do not expose these weights as a public endpoint without an independent moderation layer. Abliteration removes a direction, not a policy; some refusals survive (multi-turn re-assertion, system-prompt steering, vision-path refusals). Always report the **delta against the base model**. Re-measure on your own prompts. Whoever deploys it owns the moderation layer the original guardrails were carrying.

## Variants in this family
Hub collection: [https://huggingface.co/collections/junafinity/ornith-15-uncensored-6a896c737cf40ad660af2ebd](https://huggingface.co/collections/junafinity/ornith-15-uncensored-6a896c737cf40ad660af2ebd)


| Model | Base | Format | Precision | Notes |
|---|---|---|---|---|
| **Ornith-1.5-9B-uncensored** ← *you are here* | Ornith-1.5-9B | Safetensors (bf16) | 16-bit | Full-precision abliterated weights |
| [Ornith-1.5-9B-uncensored-MLX-8bit](https://huggingface.co/junafinity/Ornith-1.5-9B-uncensored-MLX-8bit) | Ornith-1.5-9B | MLX | 8-bit | Apple Silicon, `mlx-vlm` |
| [Ornith-1.5-9B-uncensored-GGUF-8bit](https://huggingface.co/junafinity/Ornith-1.5-9B-uncensored-GGUF-8bit) | Ornith-1.5-9B | GGUF | Q8_0 | llama.cpp |
| [Ornith-1.5-35B-A3B-uncensored-MLX-8bit](https://huggingface.co/junafinity/Ornith-1.5-35B-A3B-uncensored-MLX-8bit) | Ornith-1.5-35B-A3B | MLX | 8-bit | Apple Silicon, `mlx-vlm` |
| [Ornith-1.5-35B-A3B-uncensored-GGUF-8bit](https://huggingface.co/junafinity/Ornith-1.5-35B-A3B-uncensored-GGUF-8bit) | Ornith-1.5-35B-A3B | GGUF | Q8_0 | llama.cpp |

4-bit and 6-bit rows that previously appeared here pointed at repos that are **not published**. They were removed so this table only lists live artifacts.

## Vision & MTP preservation

**Both the vision tower and any multi-token-prediction (MTP) block are preserved.**
Abliteration is applied *only* to the residual-writing projections inside the
language-model decoder stack — `self_attn.o_proj`, `linear_attn.out_proj` and
`mlp.down_proj` (including MoE experts). The vision tower and `mtp.*` tensors are
never read and never written by the weight edit, so they carry through unchanged
by construction.

Audited at the **start and end** of the abliteration run:

| Component | Before | After | Status |
|---|---|---|---|
| **Vision tower** | 333 tensors / 456,010,480 params | 333 tensors / 456,010,480 params | ✅ **preserved** — bit-identical |
| **MTP head** | not present in base | not present | ➖ none in this lineage |

**Verification performed:**

- Tensor-name and parameter-count audit of the checkpoint before and after the run.
- SHA-256 comparison of raw tensor bytes: sampled vision-tower weights are **bit-identical** pre/post, as are all non-target language-model weights; only the intended abliteration targets differ.
- End-to-end multimodal generation on the abliterated weights (image in → description out), confirming the vision path is not merely present but functional.

> **On MTP, precisely:** the base checkpoint's `config.json` declares `mtp_num_hidden_layers: 1`, but the published weights ship **no** `mtp.*` tensors — there is no MTP block in this lineage to begin with. Nothing was removed and nothing was lost; the pipeline preserves `mtp.*` tensors wherever a checkpoint actually provides them.

## Abliteration result

| Metric | Value |
|---|---|
| Refusals on held-out harmful set | **9 → 0** / 64 |
| KL divergence from base | **0.001668** |
| Optuna trials | 100 |
| Pareto points | 4 |
| Selected trial | #90 |
| Ablation strength | 1.343 |
| Layers edited | 15–20 of 32 |
| Direction source layer | 20 |

ZeroFuse co-minimizes two objectives — remaining refusals and KL divergence from the
original model — with a multi-objective Optuna TPE search, then materializes the
selected point on the Pareto front as a direct weight edit
(`W' = W − strength · r(rᵀW)`). There is no runtime adapter and no inference-time
overhead: the result is a standard checkpoint of identical shape and speed.

The very low KL (0.001668) means the output distribution on harmless
prompts is nearly unchanged from the base model, i.e. refusal behaviour was removed
with minimal collateral effect on general capability.

## Method

1. Residual-stream activations captured on harmful vs. harmless prompt sets.
2. Refusal direction estimated by difference-of-means, with projected refinement.
3. Two-objective Optuna TPE search over source layer, layer span and strength.
4. The selected configuration orthogonalized out of the residual-writing weights.

## Usage

```python
from transformers import AutoModelForImageTextToText, AutoProcessor

model = AutoModelForImageTextToText.from_pretrained(
    "junafinity/Ornith-1.5-9B-uncensored", dtype="auto", device_map="auto"
)
processor = AutoProcessor.from_pretrained("junafinity/Ornith-1.5-9B-uncensored")
```

> Requires `transformers >= 5.12` for the `qwen3_5` architecture.

## Responsible use

Primary intended use is **red teaming and defensive cybersecurity research**. See the section of that name above.

This model has had safety guardrails **reduced or removed**. Do not expose it as a public endpoint without an independent moderation layer. You are responsible for compliance with the base model's license and acceptable-use policy, applicable law, and the terms of any platform you deploy on. Removing guardrails does not remove accountability.