base_model: ornith-ai/Ornith-1.5-9B
base_model_relation: finetune
library_name: transformers
license: mit
license_link: https://huggingface.co/ornith-ai/Ornith-1.5-9B/blob/main/LICENSE
pipeline_tag: text-generation
tags:

• abliterated
• uncensored
• heretic
• qwen3_5

Ornith-1.5-9B-Uncensored

Decensored (abliterated) build of ornith-ai/Ornith-1.5-9B.

Refusal directions were removed with Heretic, which runs a
TPE search over per-layer ablation strengths for the attention output and MLP down projections,
co-optimizing refusal rate against KL divergence from the original model. No fine-tuning or
retraining is involved, so the base model's capabilities are preserved apart from the measured
distribution shift below.

Results

• Keyword-marked refusals on 100 harmful prompts (mlabonne/harmful_behaviors test split): 85/100 -> 55/100
• KL divergence on harmless prompts: 0.0017 (lower is closer to the original)
• Search: 100 trials, exported trial 79, bf16 on an NVIDIA A40,
  base revision 98db59be66b580b0395b3dc8237b32eefcdfec22

A KL divergence this low means behaviour on ordinary prompts is essentially unchanged; the edit is
targeted at refusal behaviour.

Note on the refusal metric: Heretic's keyword scorer flags responses containing markers such as
"illegal", "harmful" or "I can't", which also appear in compliant answers that discuss sensitive
topics, so the true refusal rate is lower than 55/100. That said, the reduction is more modest than
Heretic typically achieves on dense models — this hybrid architecture appears comparatively
resistant to ablation, and the numbers are reported as measured rather than cherry-picked.

Architecture note

Ornith-1.5-9B is a Qwen3.5 hybrid: 32 layers interleaving gated DeltaNet linear-attention blocks
with full attention every 4th layer, plus a vision tower (it is multimodal), 262k context.
You need a very recent transformers (>= 5.12) - older versions fail with
unknown architecture 'qwen3_5'. Note the shipped config sets use_cache: false, which makes
generation considerably slower than typical dense models.

Usage

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model = AutoModelForImageTextToText.from_pretrained(
    "zaakirio/Ornith-1.5-9B-Uncensored", dtype="bfloat16", device_map="auto"
)
processor = AutoProcessor.from_pretrained("zaakirio/Ornith-1.5-9B-Uncensored")

messages = [{"role": "user", "content": "Explain how a Kalman filter works."}]
inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True, return_tensors="pt"
).to(model.device)
output = model.generate(**inputs, max_new_tokens=512)
print(processor.decode(output[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

Caveats

This model has had its refusal behaviour reduced. It is more likely to answer requests that the
original model declines, and it has fewer safety guardrails. You are responsible for how you use
it. Abliteration can also make a model more compliant with any framing, including incorrect
premises, so verify factual output as you would with any small model.

Inherits the MIT license
from the base model.